GoFlow: efficient transition state geometry prediction with flow matching and E(3)-equivariant neural networks

Area of Science:

• Computational Chemistry
• Machine Learning
• Chemical Reaction Dynamics

Background:

• Transition state (TS) geometries are crucial for understanding chemical reactions and kinetics.
• Current quantum chemical methods for TS computation are computationally expensive for large systems.
• Deep learning diffusion models show promise for TS generation from 2D reaction graphs but suffer from slow inference.

Purpose of the Study:

• To develop a faster and more accurate deep learning method for generating transition state geometries from 2D reaction graphs.
• To enable the practical application of machine learning-based TS generation in high-throughput computational chemistry.

Main Methods:

• Modeling TS generation as an optimal transport flow problem.
• Utilizing E(3)-equivariant flow matching with geometric tensor networks.
• Developing a novel method named GoFlow.

Main Results:

• Achieved over a hundredfold speedup in inference compared to diffusion models.
• Demonstrated improved geometric accuracy of generated TS structures.
• Enabled efficient TS generation for larger and more complex molecular systems.

Conclusions:

• GoFlow offers a significant methodological advancement in machine learning-based TS generation.
• The method's efficiency and accuracy facilitate its use in high-throughput computational chemistry.
• GoFlow brings accurate and rapid TS prediction closer to widespread adoption in chemical research.