MIT develops CRESt multimodal materials discovery copilot

MIT researchers have developed CRESt (Copilot for Real-world Experimental Scientists), a multimodal platform that integrates literature, imaging, large language models, active learning, and robotic high-throughput experimentation to optimize materials recipes and accelerate discovery.

• Main announcement/action: CRESt was used to explore more than 900 chemistries over three months and conduct 3,500 electrochemical tests, discovering an eight-element catalyst that achieved a 9.3-fold improvement in power density per dollar over pure palladium and delivered a record power density in a direct formate fuel cell while using one-fourth the precious metals of prior devices; the platform combines a liquid-handling robot, carbothermal shock system, automated electrochemical workstation, automated electron/optical microscopy, and a natural-language user interface that can include up to 20 precursor molecules and substrates.
• Background and implementation details: The approach blends literature-derived knowledge embeddings with experimental data using principal component analysis to reduce search space and Bayesian optimization for experiment selection; CRESt monitors experiments with cameras and vision-language models to detect and suggest fixes, and the work is reported in a paper published in Nature with lead authors including Ju Li, Zhen Zhang, Zhichu Ren, Chia-Wei Hsu, and Weibin Chen; human researchers remain responsible for most debugging and decision-making.