Harnessing Machine Learning to Enhance Transition State Search with Interatomic Potentials and Generative Models.

Area of Science:

• Computational Chemistry
• Chemical Reaction Mechanisms
• Machine Learning in Chemistry

Background:

• Transition state (TS) search is vital for understanding chemical reactions but computationally expensive.
• Machine learning interatomic potentials (MLIPs) and generative models offer potential acceleration for TS searches.
• Comparative performance and limitations of these machine learning (ML) approaches for TS search are not well-defined.

Purpose of the Study:

• Establish a systematic benchmarking framework to evaluate ML methods for TS search.
• Provide a standardized and application-relevant assessment of ML performance in this area.
• Compare the effectiveness of MLIPs and generative models in accelerating TS discovery.

Main Methods:

• Developed an end-to-end workflow for benchmarking TS search algorithms.
• Evaluated seven representative MLIPs and the generative model React-OT.
• Utilized a consistent graph neural network architecture for direct comparison.

Main Results:

• Pre-trained foundation MLIPs often require task-specific fine-tuning for reliable TS localization.
• Standard energy and force metrics are insufficient for predicting TS search success.
• React-OT, a generative model, frequently outperformed MLIPs with identical architectures.

Conclusions:

• Generative models show significant potential for advancing TS discovery in chemical reactions.
• A need exists for tailored evaluation criteria beyond traditional metrics for ML-based TS searches.
• This benchmark provides a foundation for developing more generalizable and reliable ML methods for reactive chemistry.