MIT researchers unveil multimodal AI model that links crystal structures to properties

In a paper published in Nature Communications, the researchers compared a version of the model trained on the multimodal dataset with a baseline trained only on text from papers. The multimodal model identified better material designs, correctly surfacing candidates already known for solid‑state lithium‑ion battery electrolytes, and learned geometric representations that text‑only models missed. Li likened the approach to AlphaFold’s linking of protein structure to function, suggesting it could become a step change for materials science.

The work 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, highlighting cross‑agency interest in accelerating materials discovery.