The work was led by Principal Researcher Sung Mook Choi of the Korea Institute of Materials Science (KIMS), in collaboration with teams headed by Professor Min Ho Seo of Pukyong National University and Professor Won Bae Kim of Pohang University of Science and Technology. Their catalyst, an ordered platinum–nickel, or PtNi, material, was developed for the hydrogen evolution reaction at the cathode of an anion exchange membrane water electrolyzer. The findings, published in Carbon Energy, offer a potential route toward longer-lasting electrolyzers that use less precious metal while maintaining the high reaction rates needed for industrial hydrogen production.

Water electrolysis separates water into hydrogen and oxygen using electricity. When that electricity comes from renewable sources such as wind or solar power, the process can produce green hydrogen without directly emitting carbon dioxide. Anion exchange membrane water electrolysis is particularly attractive because it operates under alkaline conditions and may reduce dependence on expensive platinum-group metals compared with traditional proton exchange membrane systems. However, alkaline environments make the hydrogen evolution reaction kinetically slower, meaning that highly active catalysts are required to produce hydrogen efficiently at commercially useful current densities.

Platinum is among the most effective materials for driving hydrogen evolution, but its cost and limited availability create obstacles to large-scale deployment. Alloying platinum with nickel can reduce the amount of platinum required and can also modify the catalyst’s electronic structure in ways that improve hydrogen production. The problem is that nickel is chemically less stable than platinum under operating conditions. During extended electrolysis, nickel atoms can leave the alloy and enter the surrounding electrolyte as dissolved ions or hydroxide-containing species. As nickel is removed, the catalyst’s composition, surface structure and electronic properties change, gradually reducing its ability to generate hydrogen.

The Korean research team addressed this problem by controlling not only the chemical composition of the catalyst, but also the precise arrangement of its atoms. In a conventional disordered PtNi alloy, platinum and nickel atoms occupy lattice sites in a largely random pattern. This random structure can contain configurations in which nickel is relatively weakly bound and therefore vulnerable to dissolution. The new catalyst uses an ordered intermetallic structure, in which platinum and nickel occupy well-defined positions within the crystal lattice. According to the researchers’ computational analysis, this arrangement strengthens the stabilization of nickel and raises its resistance to leaching under alkaline electrolysis conditions.

The catalyst was prepared through a two-stage process. First, platinum and nickel precursors were chemically reduced at low temperature using sodium borohydride, or NaBH4, producing a material in which the two elements were initially mixed without long-range atomic order. The powder was then heat-treated under a nitrogen atmosphere. This controlled thermal treatment gave the atoms enough mobility to rearrange into an ordered configuration while shielding the material from unwanted reactions with oxygen in the air. The resulting catalyst was deposited on an electrode and installed at the cathode, where water is converted into hydrogen through a sequence of electrochemical steps involving water molecules, electrons and adsorbed hydrogen intermediates.

The difference between the ordered and disordered materials became especially clear after durability testing. The conventional disordered catalyst lost approximately 54% of its original nickel content, indicating extensive dissolution during operation. By contrast, the ordered PtNi catalyst lost only about 9% of its nickel. This substantial reduction in leaching suggests that the crystal structure acts as an atomic-scale anchor, holding nickel within the alloy and preserving the electronic environment responsible for catalytic activity. Because the active material remains more chemically intact, the electrode can continue to promote hydrogen evolution without undergoing the rapid compositional drift that typically accelerates performance loss.

The researchers then moved beyond small-scale electrochemical measurements and tested the catalyst in a large-area three-cell stack with an active area of 64 square centimeters. This step is important because catalysts that perform well in laboratory half-cell experiments often encounter new challenges when incorporated into membrane assemblies and connected in multi-cell systems. Larger devices introduce factors such as uneven water distribution, gas management, electrical resistance, temperature gradients and fluctuations in operating conditions. Despite these practical complications, the ordered PtNi catalyst maintained stable operation for 3,000 hours, with performance degradation remaining below 2%. The result provides unusually strong evidence that atomic ordering can translate from a materials-science concept into a working electrolyzer architecture.

The advance could have consequences beyond catalyst lifetime. In commercial hydrogen facilities, frequent replacement of degraded components raises maintenance expenses, interrupts production and complicates the integration of electrolyzers with intermittent renewable power. A catalyst that retains its structure for longer periods could lower the cost of hydrogen by extending operating intervals and reducing the need for repairs. The ordered PtNi design also provides a way to reduce platinum loading without abandoning the high intrinsic activity associated with platinum. The team says the same strategy may be adaptable to other platinum–transition-metal catalysts used in fuel cells, electrolyzers and broader electrochemical energy technologies.

“Our study is significant because we used computational science to explain how atomic ordering suppresses nickel leaching under anion exchange membrane water electrolysis conditions and then experimentally validated the mechanism through detailed catalyst analysis and 3,000 hours of operation in a commercially relevant large-area three-cell stack,” said Sung Mook Choi, principal researcher and project leader at KIMS. The researchers are now working to reduce precious-metal loading further, improve the uniformity of large-area electrode manufacturing and optimize stack operating conditions. They also plan to examine how renewable-energy-driven load fluctuations affect the catalyst and to identify the detailed mechanisms responsible for degradation during longer-term operation. If those efforts succeed, locking atoms into place could become an important design principle for making green hydrogen systems more durable, affordable and ready for industrial deployment.

Subject of Research: Atomic ordering in platinum–nickel electrocatalysts for durable anion exchange membrane water electrolysis and green hydrogen production.

Article Title: Locking Ni Atoms in Ordered PtNi for Durable Hydrogen Production: From Electrocatalyst Design to Practical AEMWE Stack Validation

News Publication Date: 1-Jul-2026

Web References: Korea Institute of Materials Science (KIMS); https://doi.org/10.1002/cey2.70265

References: Choi, S. M. et al., “Locking Ni Atoms in Ordered PtNi for Durable Hydrogen Production: From Electrocatalyst Design to Practical AEMWE Stack Validation,” Carbon Energy, DOI: 10.1002/cey2.70265.

Image Credits: Korea Institute of Materials Science (KIMS)

Keywords

Green hydrogen, water electrolysis, anion exchange membrane water electrolysis, AEMWE, platinum–nickel catalyst, PtNi, atomic ordering, nickel leaching, hydrogen evolution reaction, electrocatalysis, electrolyzer durability, renewable energy, fuel cells, electrochemical energy systems

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