Extrapolation Techniques in Database Construction for Machine-Learning Potentials: Achieving Subchemical Accuracy in...

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Engineering

Background:

• Accurate prediction of catalytic reaction pathways is crucial for catalyst design.
• First-principles calculations, while accurate, are computationally expensive for large-scale screening.
• Machine-learning interatomic potentials (MLIPs) offer a way to accelerate these calculations.

Purpose of the Study:

• To present a computational workflow, CSCP-X, for constructing highly accurate MLIPs.
• To accelerate the exploration of potential energy surfaces for complex catalytic reactions.
• To achieve subchemical accuracy for predictive catalyst screening.

Main Methods:

• Developed the conformal sampling of catalytic processes enhanced with extrapolation techniques (CSCP-X) workflow.
• Constructed MLIPs with subchemical accuracy (∼0.03 eV) relative to DFT.
• Applied CSCP-X to methanol decomposition on various metal catalysts.

Main Results:

• MLIPs achieved targeted accuracy for catalyst screening with high computational efficiency.
• Required only up to two iterative active learning steps with a modest dataset increase.
• Demonstrated accurate prediction of reaction energy barriers and uncovered alternative reaction paths.

Conclusions:

• CSCP-X enables robust and efficient acceleration of novel catalytic material discovery.
• The method achieves mechanistic transferability comparable to physics-based models.
• CSCP-X is a powerful tool for high-throughput catalyst screening and design.