Transition State Searching Accelerated by Neural Network Potential

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Engineering

Background:

• Identifying transition states is crucial for designing efficient chemical processes and catalysts.
• Current computational methods for transition state identification are often resource-intensive and iterative, limiting their practical application.

Purpose of the Study:

• To develop an enhanced computational method for accurate and efficient transition state prediction.
• To benchmark different neural network potentials and transition state locating algorithms.

Main Methods:

• Integration of neural network potentials (specifically NequIP) with physical models.
• Benchmarking against various transition state locating algorithms, including the energy-weighted Climbing Image-Nudged Elastic Band (EW-CI-NEB) method.
• Evaluation of model transferability and performance improvement through active learning.

Main Results:

• Achieved highly accurate transition state predictions by combining NequIP with EW-CI-NEB.
• Demonstrated significantly superior accuracy compared to semiempirical methods.
• Showcased greatly improved efficiency over density functional theory methods.
• Validated model transferability and enhanced performance via active learning.

Conclusions:

• The integrated NequIP-EW-CI-NEB method offers a powerful and efficient approach for transition state prediction.
• This method can directly search for transition states or provide initial guesses, substantially reducing manual effort in computational chemistry.
• The developed technique holds promise for accelerating the design of chemical processes and catalysts.