2026-05-29 韓国基礎科学研究院（IBS）
Figure 1. Machine learning model for predicting a new catalyst class by integrating distinct catalyst families
The researchers developed a machine learning model that integrates data from carbon-supported single-atom catalysts and perovskite oxide catalysts. The model predicts the activity of a new material class that was not included in training: single-atom catalysts supported on perovskite oxides. The surface atomic arrangement of the catalyst is learned as image information, while the bulk structure of the oxide is learned as graph information. By combining surface-design knowledge from single-atom catalysts with bulk-structure knowledge from perovskite oxides, the model predicts the overpotential of catalysts for the alkaline oxygen evolution reaction.
＜関連情報＞
- https://www.ibs.re.kr/cop/bbs/BBSMSTR_000000000738/selectBoardArticle.do?nttId=26725&pageIndex=1&searchCnd=&searchWrd=
- https://www.nature.com/articles/s41563-026-02622-6
深層学習による異種材料触媒の発見 Cross-material catalyst discovery via deep learning
Junseok Moon,Seungwoo Yoo,Jaehyuk Shim,Sungeun Heo,Jeong Hyun Kim,Megalamane S. Bootharaju,Kug-Seung Lee,Jaeyune Ryu,Yung-Eun Sung & Taeghwan Hyeon
Nature Materials Published:28 May 2026
DOI:https://doi.org/10.1038/s41563-026-02622-6
Abstract
The discovery of catalysts is typically confined within individual material classes, limiting insight from across material types. Here we demonstrate a machine learning approach that bridges catalyst families by identifying co-descriptors derived from two experimental datasets: single-atom catalysts (SACs) on carbon and bulk perovskite oxides. This co-descriptor set, selected through automated statistical and natural-language analyses, enabled integration of distinct experimental catalyst datasets by yielding shared activity-related chemical features. The resulting unified model, the crossbreeding neural network (CBNN), enables prediction of oxygen evolution activity in a previously untrained class—SACs on perovskite oxides. The CBNN precisely predicted performance trends of experimentally synthesized catalysts by overpotential, including a multimetallic catalyst with superior activity compared with all previous candidates. Explainable machine learning further connected descriptor importance and surface atomic contributions to activity trends. These results suggest that cross-material machine learning can accelerate the discovery of high-performance catalysts beyond known design spaces.
