Accelerating Catalytic Reaction Network Exploration via Local Fine-tuning with Universal Machine Learning Interatomic Potentials

Theoretical exploration of complex catalytic reaction networks (CRNs) is limited by the trade-off between the cost of quantum mechanical calculations and the reduced accuracy of approximate methods. We introduce the LFT-CRN, an active learning framework combining pretrained universal machine learning interatomic potentials (MLIPs) with a local fine-tuning (LFT) algorithm for efficient CRN exploration. The LFT-CRN accelerates geometry optimization, transition-state search, and vibrational analysis while maintaining consistent performance across different exchange–correlation functionals and density functional theory (DFT) settings. Applied to methanol synthesis on CuZn catalysts, the LFT-CRN achieves over a 14-fold acceleration compared with the conventional DFT workflow, retaining chemical accuracy (<1 kcal/mol) for several energy metrics. Energetics and microkinetic simulation reveal that low-coordination Cu sites with moderate Zn doping maximize both Cu–Zn synergy and catalytic activity, whereas excessive Zn reduces performance. This generalizable workflow enables high-throughput CRN exploration, thereby supporting catalyst design and optimization of industrial processes.