Accelerating Multimetallic Catalyst Discovery with Robotics and Agentic AI

Abstract

The design space of catalyst materials spans composition, processing, atomistic structure, and microstructure. As materials become more complex, the dimension of this parameter space of catalyst design grows combinatorially. Conventional active learning approaches operate on a single data stream and stay decoupled from the messy reality of experiments, limiting their efficiency and reproducibility in real-world catalyst optimization. To tackle this limitation, in a recent issue of Nature, Li and coworkers developed a robotic platform, Copilot for Real-world Experimental Scientists (CRESt), which facilitates multimetallic catalyst discovery in a multiplex parameter space by combining multimodal large vision–language models (LVLMs), knowledge-assisted Bayesian optimization (BO), and robotic automation of synthesis, characterization, and electrochemical tests. Deployed on a direct formate fuel cell (DFFC) use case, CRESt efficiently explored hundreds of compositions and thousands of tests in months to deliver an octonary multimetallic electrocatalyst with excellent device-level performance at reduced noble-metal loading. In this Commentary, we highlight CRESt’s technical merits, while also outlining a forward agenda to translate systems like CRESt from proof-of-concept, bespoke demonstrations to widely adoptable, scientifically robust agentic artificial intelligence (AI) for self-driving laboratories.