MIT researchers described CLaC, a model that connects three-dimensional crystal structures with language to retrieve materials matching requested properties.
MIT News reported on October 18 that researchers developed CLaC, a multimodal AI model designed to connect crystalline materials’ three-dimensional atomic structures with natural-language descriptions. The work addresses a gap in text-based systems: many crystal structures are stored in computational databases rather than described in scientific papers, while the published literature covers some materials more heavily than others.
CLaC uses separate encoders for crystal graphs and text, aligning their representations through contrastive learning. The researchers trained it using descriptions generated from computed material properties alongside scientific literature. Their paper reports training on roughly 126,000 materials; the synthetic narratives covered about 300 times more diverse crystals than the human-written academic literature used for comparison.
In a search task, the team asked the model to find crystals suited to solid-state lithium-ion battery electrolytes. Its leading results included candidates that independent studies had already identified as promising. The researchers also reported that CLaC learned to align geometric features of crystal structures with language, while a comparison model trained only on literature text did not show the same alignment. The system could retrieve crystal structures in response to natural-language descriptions of desired properties, outperforming a baseline in the reported tests.
The approach could help materials researchers narrow large candidate pools before undertaking more expensive computational work or laboratory experiments. It is not evidence that the model has produced a ready-to-use battery material: the reported findings concern computational retrieval and representation, and candidate materials still require further validation. The work’s broader contribution is a way to make materials databases more accessible to language-guided search without relying solely on the uneven coverage of published papers.
The research was published in Nature Communications in August 2026; MIT’s October 18 article brought the work to wider attention. The team says the approach may extend across crystalline materials and hopes other materials scientists will use it. The project received support from U.S. and South Korean research agencies.