In the relentless pursuit of sustainable energy solutions, the efficient production of green hydrogen stands as a critical milestone. Central to this effort is the oxygen evolution reaction (OER), a kinetically challenging step during water electrolysis that demands substantial energy input. Improving catalyst performance for OER could revolutionize water splitting technologies, facilitating widespread clean hydrogen fuel availability. Historically, catalyst development has been constrained to individual material families, such as metal oxides, single-atom catalysts, or perovskites. This compartmentalized approach has inherently limited the exploration of hybrid material systems and cross-family synergies. However, a pioneering study from the Institute for Basic Science (IBS), led by Director HYEON Taeghwan of the Center for Nanoparticle Research, heralds a paradigm shift in catalyst discovery by leveraging artificial intelligence (AI) to integrate knowledge across diverse catalyst classes.
The team’s breakthrough centers around an innovative deep learning framework coined the Crossbreeding Neural Network (CBNN). Unlike conventional AI models trained within narrowly defined material boundaries, CBNN simultaneously assimilates data from two chemically distinct catalyst families: carbon-supported single-atom catalysts and perovskite oxide catalysts. This dual learning strategy enables CBNN to capture complementary insights—single-atom catalysts elucidate atomic-scale surface activity while perovskite oxides provide critical information on bulk crystal structure effects. The result is an unprecedented ability to predict properties of a hybrid catalyst class previously unexplored: single-atom catalysts anchored upon perovskite oxide supports.
Mechanistically, this hybrid system exploits the atomic precision of single metal atoms dispersed on catalytically active substrates, merged seamlessly with the structural and electronic versatility inherent in perovskite frameworks. Surface atomic arrangement characteristics are encoded in the model as image data, allowing the AI to interpret local atomic environments visually. Simultaneously, the bulk crystal structure of the oxide, represented by its graph information, informs the network about extended lattice periodicity and connectivity. By integrating these multidimensional data modalities, CBNN constructs a holistic representation of catalyst behavior anchored in both micro- and macro-scale structural features.
To enhance the model’s predictive rigor and interpretability, the research group implemented an automated descriptor-selection pipeline. This pipeline leverages classical statistical techniques coupled with advanced natural language processing (NLP) methods to identify physicochemical descriptors that robustly correlate with catalytic performance across both catalyst families. Key factors singled out include oxidation state, ionic radius, valence d-electron count, electronegativity, and coordination number. Such descriptors succinctly capture the electronic and geometric environment influencing OER activity and allow the AI to generalize learned relationships to novel material domains.
Experimental validation of the CBNN predictions involved synthesizing catalysts within the targeted hybrid family of single-atom-perovskite composites. Twenty-seven distinct catalysts spanning various elemental combinations were produced and rigorously tested under alkaline OER conditions. Impressively, the AI’s activity rankings for twelve of these synthesized materials precisely matched experimental measurements, confirming that the model did not simply interpolate within its training data but extrapolated genuine chemical insights to unknown material classes.
The research did not stop at single-element models; it further expanded into the complex realm of multimetallic catalysts. By computationally screening a vast candidate space of 8,008 permutations containing tungsten (W), molybdenum (Mo), ruthenium (Ru), and rhodium (Rh) single atoms embedded on a calcium–praseodymium cobalt iron oxide perovskite scaffold (Ca0.8Pr0.2Co0.8Fe0.2O3−δ, or CPCF), the AI identified optimal multimetallic compositions. These predictions were experimentally verified, with the top-performing multimetallic catalyst demonstrating superior oxygen evolution rates compared to all mono- and bi-metallic counterparts tested, as well as outperforming traditional perovskite oxides and carbon-supported single-atom catalysts.
Beyond raw catalytic activity predictions, the CBNN framework integrates explainable AI techniques to unpack the atomic-scale design principles responsible for enhanced performance. Visualization of feature importance within the model highlighted synergistic effects arising from specific neighboring metal atom configurations, revealing how electronic interactions at the atomic interface promote catalytic turnover. This level of interpretability is crucial for guiding rational catalyst design and advancing mechanistic understanding beyond black-box computational models.
Director HYEON Taeghwan eloquently summarizes the significance of this work: the AI system transcended conventional boundaries by leveraging heterogeneous datasets to explore uncharted catalyst territories rather than simply optimizing within known categories. This achievement opens the door to a paradigm wherein artificial intelligence holistically synthesizes fragmented experimental knowledge and catalyzes innovation at the interface of multiple material classes.
Looking ahead, the implications of this cross-material AI-driven discovery framework extend far beyond the oxygen evolution reaction or even catalysis. Similar integrative approaches could revolutionize materials development arenas such as battery electrode formulation, complex energy storage device design, and pharmaceutical drug discovery, where disparate heterogeneous datasets remain challenging to unify. By enabling AI to “speak” the common scientific language across domains, the potential for unveiling unanticipated material architectures and performance regimes dramatically expands.
In a landscape where the grand challenge is to design materials that reconcile often competing factors such as activity, stability, and cost, the fusion of AI with cross-class learning exemplified by CBNN offers a visionary roadmap. This study, published in the esteemed journal Nature Materials, marks a watershed moment in the journey toward generalized materials artificial intelligence capable of uncovering entirely new classes of functional materials through hybridized data-driven insight.
As the clean hydrogen economy accelerates, breakthroughs like this underscore the indispensable role of machine learning in pushing the frontiers of catalyst innovation beyond current scientific constraints. The ability to rationally engineer catalysts by bridging atomic to bulk scale phenomena unlocks unprecedented efficiency pathways essential for a sustainable energy future.
Subject of Research: Not applicable
Article Title: Cross-material catalyst discovery via deep learning
News Publication Date: 28-May-2026
Web References: 10.1038/s41563-026-02622-6
Image Credits: Institute for Basic Science
Keywords: Catalysis, Oxygen evolution reaction, Green hydrogen production, Single-atom catalysts, Perovskite oxides, Deep learning, Artificial intelligence, Multimetallic catalysts, Water electrolysis, Catalyst design, Machine learning, Nanomaterials
