(87a) Graph Neural Network Distillation for Scalable Catalytic Material Exploration

Graph neural networks (GNNs) have been successfully applied in material design, including organic molecules, inorganic crystals, and catalyst surfaces. State-of-the-art GNN models like Equiformer use deep architectures with iterative message passing layers and rich feature embeddings, enhancing accuracy. However, the computational overhead limits their use in high-throughput catalytic material screening. Exploring heterogeneous catalysts is challenging due to the complexity of adsorbate configurations, catalyst compositions, and reaction-condition-induced structural changes, all of which increase computational costs and demand faster predictions. We present MPFlow, an efficient distillation method that predicts final node representations from initial embeddings of pre-trained GNN models using a continuous velocity field. By formulating message passing as a flow matching problem, MPFlow enables direct final node predictions via single-step integration, bypassing the costly iterative process. MPFlow achieves comparable energy and force accuracy to pre-trained GNNs while significantly reducing computational complexity, making it ideal for rapid catalytic adsorption evaluation. We demonstrated MPFlow’s efficiency in funnel-type catalyst screening, expanding the search space for targeted electrochemical systems and identifying a broader range of stable and active materials for advanced electrochemical applications.