Adaptive Transition-State Refinement with Learned Equilibrium Flows.

Area of Science:

• Computational chemistry
• Chemical kinetics
• Artificial intelligence in chemistry

Background:

• Identifying transition states (TSs) is crucial for understanding chemical reactions but remains computationally challenging.
• Accurate and efficient TS identification is vital for designing chemical processes, catalysts, and pharmaceuticals.

Purpose of the Study:

• To introduce a novel generative AI approach for improving initial guesses of transition state (TS) structures.
• To enhance the accuracy, robustness, and efficiency of computational chemistry methods for TS identification.

Main Methods:

• Developed a generative AI method to refine initial guesses for TS structures.
• Integrated the AI approach with existing techniques like machine-learning models and approximate quantum methods.
• Evaluated the method's performance on TS guesses from state-of-the-art ML models and tight-binding approximations.

Main Results:

• Reduced median structural error to 0.077 Å and median absolute error in reaction barrier heights to 0.40 kcal mol⁻¹ when combined with an ML model.
• Increased the success rate of locating valid TSs by 41% and accelerated high-level quantum optimization by a factor of 3 when used with a tight-binding approximation.
• Demonstrated significant improvements in accuracy and efficiency for TS searches.

Conclusions:

• The generative AI method offers a significant advancement in accurately and efficiently identifying transition states.
• This approach has the potential to accelerate reaction mechanism discovery and aid in the development of new materials, catalysts, and pharmaceuticals.