Transferable Learning of Reaction Pathways from Geometric Priors

Area of Science:

• Computational Chemistry
• Machine Learning
• Chemical Dynamics

Background:

• Identifying minimum-energy paths (MEPs) is vital for understanding chemical reaction mechanisms.
• Traditional MEP identification methods are computationally intensive.
• Current methods often require transition-state geometries or pre-optimized paths for training.

Purpose of the Study:

• Introduce MEPIN, a scalable machine learning (ML) method for predicting MEPs.
• Develop an ML approach that does not require transition-state geometries or pre-optimized paths during training.
• Enable efficient and accurate prediction of reaction pathways.

Main Methods:

• MEPIN predicts the deviation between ground-truth MEPs and geometric interpolations.
• Utilizes a symmetry-broken equivariant neural network architecture to generate continuous reaction paths.
• Employs an energy-based objective for training, incorporating geometric priors from geodesic interpolation for efficiency gains.

Main Results:

• MEPIN accurately predicts MEPs for diverse chemical reactions, including small molecule reactions and [3 + 2] cycloadditions.
• Demonstrates accurate alignment with reference intrinsic reaction coordinates.
• Achieves significant efficiency gains compared to traditional methods.

Conclusions:

• MEPIN offers a computationally efficient and scalable approach to MEP identification.
• The method generalizes well across different reaction types.
• Provides a powerful new tool for computational chemistry and reaction mechanism studies.