Redox reactions, such as dry methane reforming (DRM) and CO2 hydrogenation to syngas, are important for converting greenhouse gases into valuable fuels and chemicals. Single-atom catalysts (SACs) and exsolution catalysts are highly active and stable in these processes due to their unique atomic-level active sites. However, understanding the atomic configurations of these active sites during catalysis remains a challenge.
In this study, we apply Graph Neural Networks (GNNs) to investigate the active sites in two key catalytic systems: single-atom platinum on ceria (Pt/CeO2) and nickel exsolution in ceria (Ni/CeO2). For Pt/CeO2, we use a transfer learning-based GNN model, fine-tuned with a minimal DFT dataset, to predict stable single atom Pt configurations and coordination motifs critical for redox reactions. This approach reduces computational costs (~1000 times faster) while maintaining ab initio accuracy in predicting catalytic structures.
For Ni exsolution, we develop a non-equilibrium-aware GNN model that simulates the exsolution configurations and captures the reaction kinetics of DRM. The model identifies the rate-limiting step in DRM and provides insights into the active site structure of exsolution catalyst, emphasizing the role of the triple phase boundary and oxygen vacancies in facilitating DRM on the Ni exsolution catalyst. These findings underline the utility of GNNs in uncovering unknown active site motifs, offering an efficient computational framework for the rational design of redox catalysts in energy conversion applications.
