A research team led by Director Hyeon Taeghwan at the Institute for Basic Science (IBS) Center for Nanoparticle Research has developed an artificial intelligence framework capable of bridging distinct catalyst families to discover new materials for green hydrogen production. Published in Nature Materials, the study introduces a method that transcends traditional limitations where catalyst discovery was confined to single material groups. The AI model successfully predicted the performance of a new class of catalysts that were never included in its training data. Green hydrogen production via water electrolysis is hindered by the slow oxygen evolution reaction (OER), which consumes excessive energy. While carbon-supported single-atom catalysts and perovskite oxides have shown promise individually, their potential is often untapped when treated separately. The IBS team engineered a deep learning model called the Crossbreeding Neural Network (CBNN) to integrate knowledge from both systems. The AI treated surface atomic arrangements of single-atom catalysts as image data and bulk structures of perovskite oxides as graph information. By combining these disparate data sources, the model learned to predict the overpotential, a key metric for catalytic efficiency, for a hybrid material: single-atom catalysts supported on perovskite oxides. To ensure accuracy, the researchers employed an automated process combining statistical analysis and natural language processing to identify critical chemical descriptors, such as oxidation state and ionic radius. Experimental validation confirmed the model's predictive power, with the AI correctly ranking the activity of 12 newly synthesized catalysts within this unexplored material family. Co-first author Moon Junseok noted that the system demonstrated genuine understanding rather than simple data memorization. Expanding the approach to multimetallic designs, the AI screened 8,008 candidates to identify a specific configuration containing tungsten, molybdenum, ruthenium, and rhodium atoms anchored on a calcium–praseodymium cobalt–iron oxide support. This multimetallic catalyst outperformed all previously studied perovskite oxides and carbon-supported single-atom catalysts, exhibiting a lower overpotential and higher turnover frequency. Beyond numerical predictions, the framework provided interpretable design principles, revealing how specific atomic environments and synergistic interactions between neighboring metal atoms enhance performance. Director Hyeon emphasized that the study demonstrates the ability to connect knowledge across chemically distinct systems to discover entirely new material classes. The researchers believe this generalizable approach could extend beyond catalysts to other fields facing similar data integration challenges, including battery technology, energy storage, and drug discovery, marking a significant step toward more versatile materials artificial intelligence.