A new artificial intelligence framework can predict how catalysts will perform by integrating knowledge from chemically distinct material families, a breakthrough that could accelerate the discovery of materials for clean energy technologies like green hydrogen production. Researchers led by Taeghwan Hyeon at the Institute for Basic Science developed the Crossbreeding Neural Network, which learned simultaneously from single-atom catalysts and perovskite oxide catalysts to predict the performance of an entirely new, unexplored material class.

What Makes This AI Approach Different From Traditional Materials Discovery?

Traditionally, machine learning models for catalyst discovery have been confined to narrow, predefined material domains. A researcher working on single-atom catalysts would build one model; a researcher studying perovskite oxides would build another. This siloed approach meant that insights from one material family rarely informed research in another, even when the underlying chemistry might share common principles.

The Crossbreeding Neural Network breaks this pattern by learning from two distinct catalyst groups simultaneously. Single-atom catalysts reveal how individual metal atoms behave on surfaces, while perovskite oxides provide data on how bulk crystal structures influence performance. By cross-referencing these data streams, the AI model predicted the performance of single-atom catalysts supported on perovskite oxides, a material combination that had never been systematically explored before.

How Did Researchers Validate the AI's Predictions?

The framework does not rely solely on computational predictions. The research team experimentally synthesized and tested 12 of the predicted catalysts within the new material family. The model correctly predicted the activity ranking of all 12 candidates, suggesting that it learned transferable relationships rather than simply memorizing the training data.

Following this validation, the team expanded the approach to screen multimetallic catalysts containing several different single-metal atoms simultaneously. The Crossbreeding Neural Network computationally screened 8,008 catalyst candidates and successfully identified a highly promising multimetallic single-atom catalyst containing tungsten, molybdenum, ruthenium, and rhodium atoms anchored on a calcium-praseodymium cobalt iron oxide perovskite support.

"When artificial intelligence learns the common language shared across different material families, it can suggest entirely new design directions beyond candidate spaces predefined by humans," noted Junseok Moon, co-first author of the study.

Junseok Moon, Co-first Author, Institute for Basic Science

What Chemical Factors Drive Catalyst Performance?

Using explainable artificial intelligence techniques, the research team visualized how specific atomic environments influence catalytic activity. This allowed them to identify key chemical factors strongly related to activity across both material families. These factors include:

• Oxidation State: The number of electrons an atom has lost or gained, which affects how readily it participates in chemical reactions.
• Ionic Radius: The size of charged atoms, which influences how they fit into crystal structures and interact with neighboring atoms.
• Valence D-Electron Count: The number of electrons in the outermost shell available for bonding, a key determinant of catalytic activity.
• Electronegativity: An atom's tendency to attract electrons, which affects its reactivity and bonding behavior.
• Coordination Number: The number of atoms or molecules bonded directly to a central atom, influencing its chemical properties.

How to Apply This AI Framework to Your Research Laboratory

For laboratory directors and materials scientists, this development offers practical advantages for improving research workflows and reducing computational overhead. Consider these implementation strategies:

• Integrate Heterogeneous Datasets: Combine experimental data from multiple material families into a single training dataset, allowing the AI model to discover cross-domain patterns that individual models would miss.
• Reduce Custom Algorithm Development: Instead of building separate machine learning models for each material class, use a unified framework that learns transferable relationships across chemically distinct systems, saving time and resources.
• Expand High-Throughput Screening Scope: Use the AI model to computationally screen thousands of candidate materials before experimental synthesis, prioritizing the most promising candidates and reducing the number of failed experiments.
• Leverage Explainable AI Techniques: Visualize which atomic properties drive performance in your specific material system, enabling chemists to understand the AI's reasoning and design better candidates manually.

What Are the Broader Implications for Materials Science?

The research suggests that similar approaches could eventually be applied to other areas involving the integration of heterogeneous datasets, including battery materials, energy-storage systems, and potentially drug discovery. The oxygen evolution reaction, a slow and energy-intensive step that occurs during water electrolysis, represents one of the central challenges in developing clean-energy technologies. By accelerating catalyst discovery, this AI framework could help make green hydrogen production more economically viable.

The findings demonstrate how AI models can transfer insights between chemically distinct material systems, potentially expanding the scope of autonomous materials discovery beyond what traditional methods allow. Rather than requiring researchers to manually identify connections between different material families, the AI learns these relationships automatically, opening new design directions that humans might not have considered.