Locating Ab Initio Transition States via Geodesic Construction on Machine-Learned Potential Energy Surfaces

Area of Science:

• Computational Chemistry
• Chemical Physics
• Materials Science

Background:

• Identifying transition states is crucial for understanding chemical reactions but computationally expensive.
• Current methods like chain-of-states often require numerous ab initio calculations and risk converging to incorrect structures.
• Potential energy surfaces (PES) are key to accurately modeling molecular behavior.

Purpose of the Study:

• To develop a novel, efficient method for obtaining high-quality transition state guesses.
• To reduce the reliance on computationally intensive ab initio calculations for transition state optimization.
• To accelerate the elucidation of complex chemical reaction networks.

Main Methods:

• Constructing geodesic paths between reactant and product structures on a machine learning potential (MLP) generated PES.
• Developing and implementing an algorithm for geodesic path optimization.
• Utilizing the eSEN-sm-cons MLP for demonstration.

Main Results:

• High-quality transition state guesses were obtained using MLP-generated geodesic paths.
• The highest-energy point on MLP geodesics required 30% fewer optimization steps compared to the ab initio freezing string method.
• The proposed approach eliminates the need for ab initio calculations in generating transition state guesses.

Conclusions:

• Geodesic path construction on ML PES offers a significant speedup for transition state optimization.
• This method provides a reliable alternative to traditional approaches, reducing computational burden.
• The approach is promising for efficient computational studies of complex chemical reaction networks.