Comprehensive exploration of graphically defined reaction spaces

Thermodynamic and kinetic properties of reactions are essential to elucidate detailed reaction mechanisms and predict reaction yields. However, limited amount of reaction databases have been built to include these quantitative chemical reaction data, especially activation energies. In this study, we utilized the concept of model reaction to comprehensively explore the reaction space and generated 177k organic reactions involving C, H, O and N atoms. These reactions are calculated at GFN2-xTB and B3LYP-D3/TZVP levels of theory and show a great coverage of the entire reaction space. A delta-learning model was then trained on this dataset, which can approach the chemical accuracy in reproducing DFT level activation energies and can even outperform DFT calculations with the introduction of a small number of G4-level computations. We believe this large quantitative chemical reaction database can accelerate progress in the development of novel machine learning models for predicting reactions properties and general machine learning force field.