Graphically-Defined Model Reactions are Extensible, Accurate, and Systematically Improvable

Achieving fast and accurate reaction prediction is central to a suite of chemical applications. Nevertheless, classic approaches based on templates or simple models are typically fast but with limited scope or accuracy, while the emerging machine learning-based models are limited in their transferability due to the lack of large reaction databases. Here, we address these limitations by formalizing the model reaction concept based on fixed-depth condensed reaction graphs that are shown to achieve a cost and accuracy balance that is applicable to many problems. The model reaction concept can be utilized to provide reliable predictions of activation energies and transition state geometries for a large range of organic reactions. In addition, using an alkane pyrolysis system as a benchmarking example, we show that the accuracy of the activation energy prediction can be further improved by adding correction terms based on the empirical Br{\o}nsted-Evans-Pokanyi (BEP) relationship. These successful applications demonstrate that the model reaction can serve as a general tool to reduce the cost associated with ab initio transition state searches.