A Machine Learning Interatomic Potential Data Set and Model for Catalysis with Local Fine-Tuning to Chemical Accuracy

Heterogenous catalysis involves complex reactions with dynamic changes in catalyst morphology, challenging the capabilities of traditional density functional theory (DFT) methods. To address this, we introduce the catalytic large atomic model (CLAM), a comprehensive data set and pretrained model for machine learning interatomic potential, specifically designed for heterogeneous catalysis. CLAM is trained on a diverse data set that includes metal and alloy slabs with adsorbates, oxides, clusters, two-dimensional materials, and small molecules, ensuring high accuracy across a wide range of catalytic systems. Additionally, we introduce a "local fine-tuning" algorithm, which substantially improves the accuracy of machine learning interatomic potentials for structural optimizations and transition state searches. This approach achieves 94% prediction accuracy within chemical accuracy thresholds (<1 kcal/mol) for adsorption energies of small adsorbates on transition metal surfaces, while demonstrating 3.4× computational acceleration relative to conventional DFT calculations. Furthermore, the method maintains comparable performance in transition state searches, delivering 81% identification accuracy (<1 kcal/mol) with a 10.1× speed-up compared to standard DFT-based CI-NEB approaches. Furthermore, the pretrained CLAM model can accurately reproduce reported dynamic catalysis phenomena via molecular dynamics simulations without additional fine-tuning.