MIT AI model links 3D crystal structures with text to find new materials

Researchers at the Massachusetts Institute of Technology have developed a multimodal AI model that combines the three-dimensional atomic structures of crystalline materials with natural language descriptions. The work, published in Nature Communications and announced on October 18, 2026, shows the system can identify better material designs than models trained on text alone.

The challenge is largely about data. Most solid materials are crystals, periodic 3D arrangements of atoms whose structures are kept in specialized file formats rather than the text that large language models learn from. The experimental literature also concentrates on a handful of popular crystal classes, which pushes AI suggestions toward those same materials.

Teaching a model to see structure

The team generated rich narratives for crystal structures, drawing on experimental papers and first-principles computational databases. Each description covers composition and structure as well as properties, processing considerations and possible applications. The model was trained with contrastive learning, comparing matching and mismatched pairs of structures and text to build a shared representation.

When the researchers compared a version trained on the synthesized descriptions with one trained only on text from papers, only the former learned to align geometric structures with language. Given a plain-language query describing a property, the model retrieved crystals with that property more successfully than a baseline.

Asked for materials suited to solid-state lithium-ion battery electrolytes, the model ranked among its top results candidates that other work had independently flagged as promising. "Structure is really at the heart of the materials science field," said Ju Li, a professor in MIT's Department of Materials Science and Engineering.

An AlphaFold for crystals

Li likens the approach to AlphaFold, the AI system that connects protein structure to function and has aided biology and drug discovery. Co-authors include first author Yang Jeong Park, Mayank Kumaran, Chia-Wei Hsu and professor Elsa Olivetti, who said she hopes the method could mark a step change for the field.

The research was supported in part by the U.S. Defense Threat Reduction Agency, the U.S. Defense Advanced Research Projects Agency and the National Research Foundation of Korea. The team believes the method applies to many crystalline materials and hopes other scientists will adopt it.