Convolutional neural networks and volcano plots for screening and predicting two-dimensional single-atom catalysts in CO2 reduction

Single-atom catalysts (SACs) emerge as pivotal frontiers for catalyzing diverse reactions, but identifying optimal combinations of active elements, supports, and coordination environments eludes traditional methods. This study integrates multi-branch convolutional neural network (CNN) models with hybrid descriptor-based activity volcano plots to screen 2D SACs for efficient CO 2 reduction reaction (CO 2 RR). Since single electronic descriptors, such as the d -band center, are insufficient to predict catalytic activity, our CNN model uses 2D electronic density of states (eDOS) to predict adsorption energies. Hybrid descriptors combining both C- and O-type CO 2 RR intermediates are used to construct volcano plots and limit potential maps for screening catalysts capable of efficient CO 2 RR. To explore the electronic scale principle governing practical CO 2 RR activity, orbital-wise eDOS occlusion and shifting analysis based on the CNN model are employed to examine the adsorption energy and, consequently, catalytic activities while varying supported single atoms, which offers a tangible framework to design efficient SACs.