Machine learning (ML) is emerging as a powerful tool in modeling potential energy surface (PES). Recently, ANI neural network potentials have been established for accurate and transferable energy prediction for organic molecules, showing orders of magnitude faster calculation speed than DFT. Current ANI models cover molecular space including rotations, clusters, non-covalent interaction but not for bond forming and breaking, which are crucial in reaction space. This work is aimed to develop a general reactive ANI potential for elementary organic reactions, further extending chemical space from molecules to reactions. A systematical generated reaction dataset is considered in this work, which contains a large number of reactions with detailed transition state information. To learn reaction PES, active learning methodologies are applied to effectively sample non-redundant data from DFT PES. A large number of reactions are organized and categorized by different reaction types with cheminformatics tools. Current results show that reactive ANI potential tends to behave well on new reactions with learnt reaction types. Predictions on new reactions with unknown reaction types are usually not good. The future plan is to develop a workflow to automatically sample data from a large number of new reactions, as well as a user-friendly interface to help user identify if the prediction of a target reaction is suitable for reactive ANI potentials.

Application of ANI deep learning in elementary chemical reactions

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.