Curated Knowledge is the Critical Path to an AI Co-Scientist for Materials Discovery

Materials discovery in industrial R\&D depends on knowledge scattered across internal reports, lab notebooks, characterization images, chemical databases, published literature, and the tacit reasoning of scientific teams. We report on three years of building and deploying a production LLM-based AI Co-Scientist for electronic materials R\&D, serving ${\sim}$700 scientists across diverse electronic material systems. We organize the system around six knowledge layers (semantic, experimental, structured, external, visual, and episodic), each unlocking a distinct capability for materials discovery. We demonstrate the system through a collaborative ideation session where 13 scientists collectively interact with our AI Co-Scientist to produce novel inhibitor chemistries for an unsolved high-temperature selectivity problem, with two candidates independently confirmed by DFT. Adoption grew 12.5$\times$ in six months once training, access, and awareness barriers were addressed. Our experience shows that curating domain knowledge, not model capability alone, is the critical path to a productive AI Co-Scientist for materials discovery.