"AI Discovers Top Catalyst"... IBS Develops Next-Generation Catalyst for Green Hydrogen [Reading Science]

Simultaneous Learning of Diverse Catalyst Data Leads to Superior Performance Validation
"AI Predicts New Material Groups"
Published in Nature Materials

Artificial intelligence (AI) has simultaneously learned the properties of different catalyst material groups and discovered a next-generation green hydrogen catalyst that outperforms existing ones. This case goes beyond simply selecting the best-performing material among existing candidates; AI has proposed entirely new catalyst groups and even validated their actual performance.

On May 28, the Institute for Basic Science (IBS) announced that the research team led by Hyeon Taeg-Hwan, Director of the IBS Center for Nanoparticle Research and Distinguished Professor at the Department of Chemical and Biological Engineering at Seoul National University, has developed a technology that integrates experimental data from different catalyst material groups into a single AI model, allowing for the prediction of new catalyst groups and the discovery of the highest-performing catalysts.

Illustration to aid understanding of the research content. Provided by the research team.

Utilizing this technology, the research team explored oxygen evolution reaction (OER) catalysts required for green hydrogen production and confirmed through actual experiments that their performance surpasses that of existing catalysts.

Water electrolysis is a technology that produces eco-friendly hydrogen by splitting water with electricity. However, the oxygen evolution reaction is slow and consumes a lot of energy, making the development of high-performance catalysts a key challenge.

The problem is that catalyst performance varies depending on numerous variables, such as constituent elements, atomic arrangement, and surface structure. Since there are so many possible material combinations, conventional approaches have largely relied on the researchers' experience and repeated experiments.

Recently, AI-based machine learning has been introduced to the field of catalyst development, but most approaches have focused on finding the optimal candidate within a single material group, such as single-atom catalysts or oxide catalysts. In real industrial settings, however, it is important to find the best catalyst across multiple material groups, not just within one.

AI Model That 'Crossbreeds' Different Catalysts Developed

The research team developed a "Crossbreeding Neural Network (CBNN)" that simultaneously learns data from single-atom catalysts and perovskite oxide catalysts.

Single-atom catalysts are those in which metal atoms are fixed individually on the catalyst surface, achieving high efficiency with a small amount of metal. Perovskite oxides are a class of materials whose properties can be adjusted by combining various metal elements.

The researchers designed the AI model to simultaneously learn the "surface information" of single-atom catalysts and the "internal structure information" of perovskite oxides within a single AI model.

Image provided by the research team to aid understanding of the study content.

Subsequently, the AI was tasked with predicting a new catalyst group it had never been trained on: catalysts with metal single atoms fixed on the surface of perovskite oxides.

As a result, the performance rankings of 12 catalysts predicted by the AI matched exactly with the results of actual synthesis and electrochemical experiments. The researchers went further, designing catalysts with multiple metal single atoms fixed together, and discovered a top-performing material that surpasses existing catalysts.

"Potential for Expansion to Battery and Drug Development"

The core of this research is that the AI was designed not only to provide "black box predictions" of results but also to explain which factors contribute to improved catalyst performance.

Director Hyeon Taeg-Hwan commented, "This study demonstrates that by combining knowledge from different catalyst groups, we can find the best-performing catalysts within a broader material space, moving beyond the conventional approach of searching for the best within a single group. This approach can be extended to various fields requiring complex material exploration, such as batteries, energy materials, and drug development."

Co-first author Moon Junseok, a combined master's and doctoral student at the IBS Center for Nanoparticle Research, explained, "If AI learns the common language of multiple material groups, it can propose new design directions beyond the candidate groups predetermined by humans. This is an important starting point toward universal material AI."

The results of this study were published online in the international journal Nature Materials on May 28.

This content was produced with the assistance of AI translation services.