One of the goals of ‘AI for Science’ is to discover customized materials through real-world experiments. Pioneering advances have been made in computational predictions and the automation of materials synthesis1–7. Yet most materials experimentation remains constrained to using unimodal active learning approaches, relying on a single data stream. The potential of artificial intelligence to interpret experimental complexity remains largely untapped8,9. Here we present Copilot for Real-world Experimental Scientists (CRESt), a platform that integrates large multimodal models (incorporating chemical compositions, text embeddings and microstructural images) with knowledge-assisted Bayesian optimization and robotic automation. CRESt uses knowledge-embedding-based search space reduction and adaptive exploration–exploitation strategy to accelerate materials design, high-throughput synthesis and characterization, and electrochemical performance optimization. CRESt enables monitoring with cameras and the generation of vision-language-model-driven hypotheses to diagnose and correct experimental anomalies. Applied to electrochemical formate oxidation, CRESt explored more than 900 catalyst chemistries and 3,500 electrochemical tests within 3 months, identifying a state-of-the-art catalyst in the octonary chemical space (Pd–Pt–Cu–Au–Ir–Ce–Nb–Cr) that exhibits a 9.3-fold improvement in cost-specific performance.
