Skip to content
Neural Daily

ResearchNews

MIT model gives AI a 3D view of crystal materials

MIT researchers built a multimodal model that reads the 3D atomic structure of crystalline materials together with natural-language descriptions. According to the work in Nature Communications, it finds better material candidates than models trained on text alone.

By: Mikola Máté Patrik · · 2 min read

GitHub repository of CLaC, the MIT contrastive language-crystal model, with code and data folders
Image: GitHub · Screenshot · Screenshot of public repository page · Source

The problem

More than 90% of solid materials are crystalline, but their structures are stored in specialised formats rather than the text that language models learn from (MIT News). The literature also skews toward a few popular crystal classes, which pulls AI outputs the same way.

What the researchers did

The model, CLaC, uses contrastive learning to align 3D crystal graphs with natural-language descriptions (Nature Communications). The key is the data: instead of relying only on paper text, the team used a language model to generate synthetic “narratives” for crystals from experimental papers and computational databases. These cover composition and structure as well as properties, processing considerations and potential applications.

In comparison, the version trained on synthetic descriptions learned to align geometry with text, while the one trained on paper text alone did not.

What it is good for

You can query the model in natural language for a property and it retrieves crystalline materials that have it. Asked for solid-state lithium-ion battery electrolytes, its top results included materials that others had independently identified as promising. Lead researcher Ju Li compares the approach to AlphaFold, which links protein structure to function.

Can you try it

The code and synthetic dataset are public on GitHub (GitHub). The paper was published on August 26; MIT presented it to a wider audience on October 18. Useful for materials and battery researchers, though practical breakthroughs still need lab confirmation.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles