Machine Learning Identification of Active Sites in Graphite-Conjugated Catalysts

Graphite-conjugated catalysts (GCCs) are a promising class of materials that combine many of the advantages of heterogenous and homogeneous catalysts. In particular, GCCs containing an aryl-pyridinium active site appear to be effective nonmetal catalysts for the oxygen reduction reaction (ORR). In this study, we analyzed both structural and electronic properties of a dataset of molecules containing nitrogen atoms embedded in aromatic molecules in order to understand which properties enable a particular site to bind O2, which is a necessary step for the initiation of ORR. We found that carbon atoms ortho or para to nitrogen and at the edge of aromatic systems are especially likely to be active. Using both structural and electronic features to describe the individual atoms in each catalyst, we trained machine learning models capable of identifying catalyst active sites. Although permutation importance of the features used to train these models indicates that several key electronic features have the greatest impact on model performance, the model trained only on structural features still proved effective in simulated catalyst discovery scenarios where the objective is affected more by false positives than false negatives.