'AI Hypothesizes And Reasons Like A Scientist' — High-Performance Catalyst Design Using Multiple AI Agents

Researchers have developed MAESTRO (which stands for Multi-Agent-based Electrocatalyst Search Through Reasoning and Optimization), a multi-agent framework in which large language models (LLMs) design catalyst structures on their own and improve performance by learning from their own previous design results.

^*Multi-agent artificial intelligence: A system in which multiple AI models, each assigned a different role, communicate and collaborate like a single team to solve complex problems.

This research was conducted jointly by a team led by Professor Back Seo-in of the KU-KIST Graduate School of Converging Science and Technology/Department of Integrative Energy Engineering at KU (President Kim Dong-One), alongside research teams led by Professors Victor Fung and Guoxiang Hu of the Georgia Institute of Technology.

AI is widely employed to rapidly screen promising materials from vast pools of catalyst candidates. However, most existing methods rely heavily on pre-built training data and prediction models. Even when a good material is identified, the AI models are unable to clearly explain why that material performs well.

In MAESTRO, which was developed by the research team, multiple LLM agents collaborate to design catalysts. The Design Agent analyzes the catalyst structure and proposes hypotheses and structural modifications for improving performance, while the Reflect Agent analyzes how the structural change affects performance. The Summary Agent then accumulates this experience as a design history. The team also introduced an exploration–exploitation strategy that broadly explores various structures in the early stage and subsequently conducts optimization focusing on the most promising regions based on the accumulated information.

MAESTRO was validated using single-atom catalysts (SACs) for the oxygen reduction reaction (ORR), with the agents finding, through iterative design and feedback, a way to selectively tune the binding energy of specific intermediates, leading to the discovery of highly active catalyst structures that overcome the scaling relations, a chronic limitation of conventional catalyst research.

^*Scaling relations: A phenomenon in catalytic reactions where the binding energies of various reaction intermediates are interdependent, acting as a primary barrier preventing catalytic performance from exceeding certain thresholds.

This study demonstrates that LLM agents can iteratively carry out the research process of a scientist, including hypothesis formulation, structural design, calculation, interpretation of results, and accumulation of experience. Notably, through in-context learning, which accumulates prior experience as context without any separate retraining of the model, the agents improved their catalyst design strategy and presented it as chemical principles that humans could understand.

Professor Back Seo-in said, “If AI agents, simulations, and experiments are seamlessly combined in the future, this approach could develop into an autonomous research system that works with human researchers to formulate and verify hypotheses and uncover new principles for materials design.”