Can multiple LLM agents learn electrocatalyst design rules instead of only ranking candidates?

The electrochemical problem

Machine-learning catalyst screening usually starts from a predefined candidate space and a predefined target. It can rank candidates efficiently, but the search strategy itself often carries little chemical reasoning.

For oxygen reduction, this limitation is especially relevant because improving one adsorbed intermediate often shifts the others through scaling relations. Escaping those correlations requires proposing structural changes for a physical reason rather than only selecting the next point numerically.

What the researchers did

The authors built MAESTRO, a multiagent framework with design, reflection, summary, and exploration roles. An LLM agent proposes structural changes to single-atom catalysts, while a machine-learning force field rapidly evaluates ORR overpotential and dissolution potential.

The loop records previous candidates and outcomes so that later proposals can use the accumulated design history as in-context learning. The study compared this reasoning-driven strategy with alternative search approaches, including Bayesian optimization.

What they found

Across four discovery strategies, the study analyzed 2,017 unique catalysts from 4,000 generated structures. Conventional ORR scaling relations imply a theoretical minimum overpotential of about 0.36 V when all intermediates move together.

MAESTRO generated candidates that departed from those relations, including structures where local interactions selectively stabilized intermediates. The authors attribute this behavior to design history and iterative feedback rather than to fixed background knowledge in the LLM.

Why this matters

The notable step is that the language model is not asked to predict a catalytic number directly. It is placed in a closed reasoning loop where proposed chemical modifications are evaluated, criticized, summarized, and reused.

That architecture points toward a different role for generative AI in electrochemistry: a system that proposes physically motivated experiments or structures and learns from the consequences of its own proposals.

One important caution

The discovered catalysts and scaling-relation breaks are computational results. The loop depends on the fidelity of the machine-learning force-field surrogate and the assumptions used to calculate activity and stability.

No experimental synthesis or electrochemical validation is reported for the new candidates, so the work demonstrates an autonomous design strategy rather than experimentally established catalyst performance.