Reactive machine learning potential for accelerating transition state search in organic synthesis

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Organic Synthesis

Background:

• Understanding reaction kinetics is crucial for organic synthesis.
• Traditional quantum chemistry methods for transition state searches are computationally intensive.
• Accurate prediction of energy barriers is essential for reaction optimization.

Purpose of the Study:

• To develop a fast and accurate machine learning model for transition state optimization and energy barrier prediction.
• To enable efficient studies of reaction kinetics in multi-element organic synthesis.
• To overcome the computational cost limitations of traditional quantum chemistry methods.

Main Methods:

• Developed DeePEST-OS, a reactive machine learning potential.
• Integrated physical priors from semi-empirical quantum chemistry.
• Utilized equivariant message passing networks for potential energy surface prediction.
• Trained on ~75,000 reactions generated via a low-cost data strategy.

Main Results:

• DeePEST-OS predicts potential energy surfaces ~10,000 times faster than quantum chemistry.
• Achieved high accuracy in transition state geometry (0.12 Å RMSD) and energy barriers (0.60 kcal/mol MAE) on unseen reactions.
• Model spans ten chemical elements.

Conclusions:

• DeePEST-OS significantly accelerates transition state optimization and energy barrier prediction.
• The model enables practical applications such as conformer screening and retrosynthesis barrier prediction.
• DeePEST-OS is a powerful tool for advancing reaction kinetics studies in organic synthesis.