MIT Researchers Build AI That Reads 3D Crystals, Improving Searches Across 90% of Solid Materials
Updated
Updated · MIT News · Oct 18
MIT Researchers Build AI That Reads 3D Crystals, Improving Searches Across 90% of Solid Materials
1 articles · Updated · MIT News · Oct 18
Summary
A new MIT multimodal model links 3D atomic structures with natural-language descriptions, letting AI evaluate crystals beyond text-only literature and better match materials to requested properties.
The system was built to address two bottlenecks: crystal data often sit in specialized computational formats rather than text, and published experiments are skewed toward a few heavily studied material classes.
In tests reported in Nature Communications, the model outperformed a text-only baseline and, for solid-state lithium-ion battery electrolytes, surfaced top candidates that had already been identified independently as promising.
More than 90% of solid materials are crystalline, and the researchers say the approach could become a broad tool for linking structure, processing and performance across materials science.