Transition1x - a dataset for building generalizable reactive machine learning potentials.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning (ML) models show promise in molecular dynamics but struggle as surrogate potentials for reaction barrier searches.
• Existing ML training datasets predominantly feature molecular configurations near equilibrium, limiting their applicability to reactive systems.

Purpose of the Study:

• To introduce the Transition1x dataset, a comprehensive resource for training ML models on configurations relevant to chemical reaction pathways.
• To address the limitations of current ML models in accurately predicting transition states and energies.

Main Methods:

• Generated 9.6 million Density Functional Theory (DFT) calculations of forces and energies for molecular configurations around reaction pathways.
• Utilized the Nudged Elastic Band (NEB) method with DFT for 10,000 organic reactions, capturing intermediate configurations.
• Trained equivariant graph message-passing neural network models on the Transition1x dataset.

Main Results:

• Demonstrated that ML models trained solely on equilibrium data fail to capture essential transition state features.
• Showcased the effectiveness of models trained on Transition1x in learning features relevant to reaction pathways.
• Cross-validated models on established datasets like ANI1x and QM9, highlighting performance differences.

Conclusions:

• The Transition1x dataset is a crucial benchmark for advancing ML force fields beyond equilibrium configurations.
• Next-generation ML models require diverse datasets that include non-equilibrium and reactive system data for broader applicability.
• This work represents a significant step towards developing more robust ML potentials for chemical reaction studies.