AI
AI agent automates discovery of new catalysts for clean energy reactions
An LLM-powered system called Catalyst-Agent has learned to autonomously screen and test materials for electrochemical applications, identifying promising new candidates for CO2 and nitrogen reduction.
Researchers have developed an artificial intelligence agent that can independently discover new catalyst materials for critical electrochemical reactions—a process that has traditionally required extensive manual work from chemists.
The automation challenge#
Finding suitable catalysts for reactions like oxygen reduction (ORR), nitrogen reduction (NRR), and carbon dioxide reduction (CO2RR) is central to advances in clean energy and chemical synthesis. While machine-learning tools have dramatically sped up the individual computations needed to evaluate candidate materials, the broader screening process remains bottlenecked by human decision-making. Scientists must select which candidates to test, prepare computational models of their structures, determine where molecules will bind, troubleshoot failed predictions, and decide which directions to pursue next. This iterative, intuitive work cannot be easily replaced by algorithms.
How the agent works#
The new system, called Catalyst-Agent, uses a large language model operating as a Model Context Protocol (MCP) server to coordinate all aspects of materials discovery. It searches public databases through the OPTIMADE standard, constructs atomic slab models, calculates binding energies using Meta FAIRchem's UMA machine-learning potential within the AdsorbML framework, scores candidates against reaction-specific chemical descriptors, and automatically modifies materials that come close to meeting desired criteria.
In tests across three reaction types, Catalyst-Agent achieved strong results, finding suitable candidates in an average of 1.40–3.41 trials per success. The system identified five new CO2RR materials—Sn₃Sc, Sn₃Y, Tl₃La, Pb₃Y, and In₃Y—that had not been previously reported for this application. Density functional theory calculations confirmed the validity of screening predictions for sample NRR and CO2RR compounds.
Why intelligent selection matters#
A critical advantage emerged when the researchers tested how much the agent's chemical reasoning contributed to its success. Replacing the agent's informed candidate selection with random screening collapsed performance dramatically: success rates dropped to 13.3% for ORR, 16.7% for NRR, and fell to 0% for CO2RR. This shows that the improvements came not from evaluating more materials through brute force, but from the agent's ability to apply chemical intuition and learn from feedback.
