Retraining and evaluating EquiformerV2 machine-learning potentials for global optimization of Mo2C nanoclusters

Nanoclusters enriched in coordinatively unsaturated metal sites usually exhibit distinct geometric and electronic properties and outperform bulk catalysts. However, identifying their atomic structures remains challenging due to vast configurational spaces to explore. In this work, we coupled genetic algorithm (GA) with retrained machine-learning potential (MLP) to accelerate configurational sampling and global minima optimization. Mo₂C nanoclusters in the gas phase and on a graphene support were used as a prototypical model system. The original and fine-tuned EquiformerV2 MLPs using Mo2C-nanoncluster DFT data were assessed for GA-MLP workflow using Kendall’s τ as a ranking metric. EquiformerV2 fine-tuned using targeted DFT data can accelerate GA-driven globle configuration exploration, provided that model evaluation focuses on correct energetic ordering (Kendall’s τ).