AI model trained on two catalyst families predicts performance of a third it has never seen

AI Model Predicts Catalyst Performance Across Unrelated Material Families

Researchers have developed a deep learning system that discovers new catalysts by combining knowledge from chemically distinct material families, according to a study published in Nature Materials. The work addresses a persistent limitation in automated materials discovery: machine learning models have traditionally been confined to narrow, predefined material domains.

The research team, led by Taeghwan Hyeon at the Institute for Basic Science, created a model called the Crossbreeding Neural Network. It learned simultaneously from two unrelated catalyst groups: single-atom catalysts on carbon materials and perovskite oxide catalysts. By cross-referencing these datasets, the model predicted the performance of a completely new material class that researchers had never studied: single-atom catalysts on perovskite oxides.

Validation through experimental testing

The team synthesized and tested 12 catalysts predicted by the model. The system correctly ranked all 12 by activity level, suggesting it learned transferable relationships rather than memorizing training data.

The researchers then expanded the approach to screen multimetallic catalysts containing multiple single-metal atoms. The Crossbreeding Neural Network computationally evaluated 8,008 candidates and identified a promising multimetallic catalyst containing tungsten, molybdenum, ruthenium, and rhodium atoms anchored on a calcium-praseodymium cobalt iron oxide perovskite support.

Practical implications for research facilities

For laboratory directors managing research operations, this approach could streamline high-throughput screening workflows and reduce the cost of custom algorithm development. Rather than building separate models for each material family, AI systems could transfer insights across chemically distinct domains.

Using explainable AI techniques, the team identified five chemical factors that strongly predicted activity across both material families: oxidation state, ionic radius, valence d-electron count, electronegativity, and coordination number. Understanding these factors allows researchers to design new materials based on principles learned from unrelated systems.

The researchers suggest similar approaches could apply to battery materials, energy-storage systems, and drug discovery-any field involving heterogeneous datasets across multiple domains.

Researchers interested in AI applications for materials science and laboratory automation may benefit from AI for Science & Research training resources.