In principle AI could make material discovery as easy as ordering dinner: scientists state what they need and a large language model returns a perfectly designed crystal. In practice most models only see text, while crystal structures are stored as 3D atomic coordinates, so performance lags. MIT researchers built a multimodal AI that fuses 3D atomic structures with natural‑language descriptions, extracting composition, geometry, properties, processing hints and potential applications from papers and first‑principles databases. The system outperformed text‑only models when asked to find solid‑state lithium‑ion battery electrolytes, surfacing candidates already identified by independent studies. Using contrastive learning, the model aligns visual geometry with textual narratives, much like the brain aligns vision and language. After training, the version that learned from AI‑generated narratives could map structures to descriptions, whereas the text‑only version could not. The model also accepts property‑based natural language queries and retrieves matching crystals better than a baseline. Researchers liken it to AlphaFold, hoping it becomes a step‑change tool that lets scientists see three‑dimensional structures and link them to function. Review