Discovering new catalysts is one of the central challenges in developing clean-energy technologies such as green hydrogen production. Yet catalyst discovery has traditionally remained confined within individual material families, limiting researchers’ ability to transfer knowledge across chemically distinct systems.
A research team led by Director HYEON Taeghwan of the Center for Nanoparticle Research within the Institute for Basic Science (IBS) has developed an artificial intelligence (AI) framework that discovers catalysts in a fundamentally new way — by combining knowledge across different catalyst families.
One of the biggest challenges in green hydrogen production is the oxygen evolution reaction (OER), a reaction that occurs during water electrolysis. Although water electrolysis can produce hydrogen without direct carbon emissions, the oxygen evolution step is slow and requires large amounts of energy. Developing better catalysts is therefore essential for improving the efficiency of green hydrogen production.
Traditionally, catalyst discovery has largely been performed within a single material family. Researchers would search for the best catalyst among oxide catalysts, metal catalysts, or single-atom catalysts separately. However, the best catalyst may emerge by combining strengths from multiple material families rather than optimizing only one group.
“The ultimate goal is not simply to find the best catalyst within one category,” said Director HYEON Taeghwan. “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 achieve this, the researchers developed a deep learning model called the Crossbreeding Neural Network (CBNN). The AI learned simultaneously from two different catalyst groups: single-atom catalysts supported on carbon materials and perovskite oxide catalysts.
These two catalyst families provide different kinds of information. Single-atom catalysts help reveal how individual metal atoms behave on catalyst surfaces, while perovskite oxides provide information about how bulk crystal structures influence catalytic performance.
By combining these two sources of knowledge, the AI was able to predict the performance of a completely new catalyst family that it had never previously encountered — single-atom catalysts supported on perovskite oxides.
In this hybrid catalyst system, individual metal atoms are anchored onto the surfaces of perovskite oxide particles, combining the advantages of both surface engineering and bulk crystal design.
To improve prediction accuracy, the researchers also developed an automated descriptor-selection process combining statistical analysis and natural language processing (NLP). The AI identified several key chemical factors strongly related to catalytic activity across both catalyst families, including oxidation state, ionic radius, valence d-electron count, electronegativity, and coordination number.
The researchers then experimentally synthesized and tested the AI-predicted catalysts. Remarkably, the AI correctly predicted the activity ranking of 12 tested catalysts within this previously unexplored material family.
“This result shows that the AI did not simply memorize existing data,” said co-first author MOON Junseok. “The model was able to judge which catalysts would perform better, even in a completely new material family it had never seen before.”
The researchers next expanded the approach to multimetallic catalysts containing several different single-metal atoms simultaneously. The AI computationally screened 8,008 catalyst candidates and identified the most promising structure: a multimetallic single-atom catalyst containing tungsten (W), molybdenum (Mo), ruthenium (Ru), and rhodium (Rh) atoms anchored on a calcium–praseodymium cobalt iron oxide perovskite support (Ca0.8Pr0.2Co0.8Fe0.2O3−δ, abbreviated CPCF).
Experimental validation confirmed that this catalyst outperformed previously studied perovskite oxide catalysts, carbon-supported single-atom catalysts, and all monometallic catalysts synthesized in this study.
Importantly, the AI model also provided interpretable design principles rather than only numerical predictions. The researchers used explainable AI techniques to visualize how specific atomic environments influence catalytic activity and identified synergistic interactions among neighboring metal atoms that enhance oxygen evolution performance.
“The significance of this work is that AI did not simply select candidates within a predefined material group,” said Director HYEON Taeghwan. “Instead, it connected knowledge across different catalyst families and used that knowledge to predict a completely new material class.”
The researchers say this framework could extend beyond catalyst discovery into broader areas such as batteries, energy storage materials, and drug discovery, where integrating heterogeneous experimental datasets remains a major challenge.
“When AI learns the common language shared across different material families, it can suggest entirely new design directions beyond candidate spaces predefined by humans,” said MOON Junseok. “This work represents an important step toward more generalizable materials AI.”
The findings were published in Nature Materials.