MIT researchers and collaborators built CLaC to let materials scientists search crystal databases with natural-language descriptions of desired properties.
Crystalline materials are made of atoms arranged in repeating three-dimensional patterns. Researchers often search materials databases by composition or use specialized tools to compare structures. CLaC, short for Contrastive Language-Crystal Pretraining, aims to connect those structures with descriptions of what a material does or could be used for.
How CLaC finds crystal structures from text queries
The system encodes each crystal as a graph that represents its atoms and their connections. A separate text encoder represents a description. During training, contrastive learning brings the paired structure and description closer together in a shared mathematical space, while unrelated pairs remain farther apart. That lets a text query retrieve crystal records whose learned representations are similar.
To create the text side of its training data, the researchers used language models to produce narratives conditioned on computed materials properties. They paired these with 126,000 materials from the Materials Project database. The researchers report that these property-based descriptions helped CLaC match text with crystal structures; the literature-derived text they tested did not produce the same match.
The paper's zero-shot retrieval tests ask the model to find crystals from natural-language descriptions without task-specific fine-tuning for each query. MIT News describes a query about solid-state lithium-ion battery electrolytes returning crystal matches from a materials database. That is database retrieval; it does not show that a candidate was synthesized or tested in a working battery.
Limits on CLaC’s results and public materials
CLaC applies to crystalline materials, and the paper evaluates retrieval and screening rather than experimental performance. Real materials can also include defects or more complex structures that the current model does not cover. A language match is therefore a lead for researchers to examine, not proof that a material has the requested properties in practice.
The authors have released the code and synthetic narrative dataset. Their repository says copyright restrictions prevent it from releasing the complete literature-derived text corpus and the corresponding pretrained weights, limiting what others can reproduce from the public materials alone.
Sources
- MIT News: Giving AI a “crystal clear” view of the materials world (highlighted Oct. 18, 2026)
- Nature Communications: Scaling multimodal materials representation learning with synthetic narratives (paper published Aug. 26, 2026; version of record Sept. 28, 2026)
- Authors’ CLaC code and data notes
- Illustrative image: ARTE via Wikimedia Commons, public domain. The diagram is not a CLaC output.
