Inside Seoul's Institute for Basic Science, researchers led by Director Hyeon Taeghwan have cracked open a longstanding puzzle in clean energy: how to leap across the boundaries between different families of catalysts to discover materials nobody has seen before. Their breakthrough, published in Nature Materials, uses artificial intelligence to combine knowledge from two chemically distinct catalyst worlds—and the result could accelerate the path toward cheaper, faster green hydrogen production.
Green hydrogen promises clean energy without carbon emissions, but there's a stubborn bottleneck. When water is split through electrolysis to produce hydrogen, the oxygen evolution reaction—the flip side of the process—demands enormous energy and moves frustratingly slowly. Better catalysts are essential to speed it up, yet for decades, catalyst discovery has been trapped inside narrow silos. Researchers would optimize oxide catalysts, or metal catalysts, or single-atom catalysts, each working in isolation. The best solution might exist in a place nobody was looking: at the intersection of multiple material families.
The team's solution is elegant. They built a deep learning model called the Crossbreeding Neural Network that simultaneously learned from two different catalyst groups: single-atom catalysts on carbon materials and perovskite oxide catalysts. These families speak different chemical languages. Single-atom catalysts reveal how individual metal atoms behave at surfaces; perovskite oxides show how bulk crystal structures drive performance. By training the AI on both, it could predict something entirely new: single-atom catalysts anchored onto perovskite oxide surfaces—a hybrid class that had never been in the training data.
"The ultimate goal is not simply to find the best catalyst within one category," Director Hyeon explained. "What researchers truly want is to identify the best catalyst across all possible material systems. We wanted to demonstrate that AI can connect knowledge from different catalyst families and use it to discover entirely new catalyst classes."
To sharpen predictions, the researchers developed an automated process that combined statistical analysis and natural language processing to identify the key chemical factors that influence catalytic activity across both families: oxidation state, ionic radius, valence d-electron count, electronegativity, and coordination number. These descriptors became the AI's compass.
The model then predicted a specific multimetallic catalyst composed of W, Mo, Ru, and Rh single atoms anchored onto the surface of Ca0.8Pr0.2Co0.8Fe0.2O3−δ perovskite oxide. The researchers synthesized and tested it experimentally. The results matched the AI's predictions: the new catalyst showed lower overpotential than previously studied perovskite oxides and higher turnover frequency than carbon-supported single-atom catalysts. Among all single-atom catalysts on perovskite oxides tested, it performed best.
What matters most is what this means for green hydrogen's future. The research proves that AI can act as a bridge between isolated domains of materials science, unlocking catalyst designs that traditional, compartmentalized research would never find. As the world races to make hydrogen production more efficient and affordable, this Seoul-based discovery offers a blueprint: sometimes the breakthrough you need isn't hiding within the system you know—it lives at the boundary where two systems meet.