An Active Learning Algorithm for Identifying Transition States on a Potential Energy Surface.

Area of Science:

• Computational Chemistry
• Chemical Physics
• Materials Science

Background:

• Mapping reaction pathways and identifying transition states (TSs) are crucial for understanding chemical reaction mechanisms.
• The standard nudged elastic band (NEB) method is effective but computationally expensive for large systems due to repeated energy and force calculations.

Purpose of the Study:

• To develop an efficient active learning algorithm, AL-NEB, for faster convergence to transition states.
• To reduce the computational cost associated with finding transition states in complex chemical systems.

Main Methods:

• An active learning algorithm (AL-NEB) was developed, integrating with the nudged elastic band method.
• The algorithm constructs a surrogate potential energy surface (PES) and uses a two-phase active learning strategy (Exploration-Exploitation and Renunciation).
• The method was tested on various systems, including 2D potentials, HCN isomerization, keto-enol tautomerization, and high-dimensional heptamer island diffusion.

Main Results:

• AL-NEB successfully located the exact transition states for all tested systems.
• The algorithm achieved convergence with an order-of-magnitude fewer force evaluations compared to the standard NEB method.
• Demonstrated scalability and efficiency for systems up to 525 degrees of freedom.

Conclusions:

• AL-NEB offers a significant improvement in efficiency and scalability for finding transition states.
• The active learning approach reduces computational burden, making complex reaction pathway mapping more feasible.
• This method holds promise for accelerating computational studies in various chemical and materials science domains.