Data-Driven Synthesis Prediction of Single-Atom Catalysts for Decarbonization and Pollution Abatement Using a Transformer-Based Framework

The practical deployment of single-atom catalysts (SACs) still faces significant challenges due to the complex, non-linear relationship between bulk synthesis parameters and the resulting atomic-scale active sites, which often requires extensive empirical optimization. We develop SACFormer, a multi-task deep learning framework, to predict the active-site configurations of graphitic carbon nitride (g-C3N4)-supported SACs directly from reactant SMILES and synthesis conditions. Trained on a manually curated dataset of 191 synthesis entries and enhanced by SMILES enumeration, SACFormer achieves a joint prediction accuracy of 87.64% for the metal center, coordinating element, and coordination number. Furthermore, we perform model uncertainty quantification to investigate the underlying mechanisms. Our analysis identifies a temperature window (450–600 °C) that favors the formation of active metastable sites, providing a strategic “transition area” regime that can inform catalyst engineering for pollution abatement (e.g., advanced oxidation processes). Conversely, our analysis indicates that, for decarbonization applications (e.g., CO2 reduction) that require specific selectivity, high-temperature annealing (>650 °C) or macrocyclic precursors are predicted to favor thermodynamic stability and stabilize the metal centers by retaining the coordination environment of the precursor. These results offer data-driven synthesis guidelines and predictive phase diagrams, highlighting how forward prediction frameworks can clarify synthesis–structure relationships.