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

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Engineering

Background:

• Heterogeneous catalysis involves complex reactions with dynamic catalyst changes.
• Traditional density functional theory (DFT) methods struggle with these complexities.

Purpose of the Study:

• Introduce the Catalytic Large Atomic Model (CLAM) for machine learning interatomic potentials in heterogeneous catalysis.
• Enhance accuracy and efficiency of catalytic simulations.

Main Methods:

• Trained CLAM on a diverse dataset including metal/alloy slabs, oxides, clusters, 2D materials, and small molecules.
• Developed a 'local fine-tuning' algorithm to improve ML interatomic potentials for structural optimizations and transition state searches.
• Utilized molecular dynamics simulations to assess CLAM's ability to reproduce dynamic catalysis phenomena.

Main Results:

• Achieved 94% prediction accuracy for adsorption energies on transition metal surfaces within chemical accuracy thresholds.
• Demonstrated 3.4x computational acceleration compared to DFT.
• Showcased 81% accuracy in transition state searches with a 10.1x speed-up over DFT-based CI-NEB.
• Successfully reproduced dynamic catalysis phenomena using molecular dynamics simulations without additional fine-tuning.

Conclusions:

• CLAM provides a highly accurate and efficient machine learning approach for heterogeneous catalysis.
• The 'local fine-tuning' algorithm significantly enhances the predictive power of ML interatomic potentials.
• CLAM shows promise for accelerating catalyst discovery and understanding dynamic catalytic processes.