MIT AI model learns to match crystal structures with plain-English descriptions

Most AI language systems learn from text. The atomic layouts of crystals, which make up more than 90 percent of Earth's solid materials, are stored in specialized 3D formats that those systems cannot easily read.

A team led by MIT materials science professor Ju Li tried to bridge that gap. Their results appear in Nature Communications, MIT News reported Oct. 18.

Building the bridge

The researchers used AI to write short descriptions for crystal structures drawn from research papers and computational databases. Each covered a material's makeup, structure, properties, processing and possible uses.

They then trained a model to pair each 3D structure with its matching description and to push mismatched pairs apart, a technique called contrastive learning.

What it can do

• Users can type a request for a property and get back candidate crystals.
• Asked for solid electrolytes for lithium-ion batteries, the model's top picks included materials other researchers had already flagged as promising.
• A version trained only on text from papers failed to link structure and language; the version trained on the generated descriptions succeeded.

The team says the work also helps with a known problem: published research focuses on a few popular crystal families, which can skew AI tools trained on it.

What's missing

The MIT article gives no benchmark scores or dataset sizes, and it does not say how accurate the machine-written descriptions are. The battery results matched known candidates rather than revealing tested new materials. Li describes the work as a step toward, not the arrival of, tools that can design materials on request.

Collaborators included MIT professor Elsa Olivetti. Funding came partly from U.S. defense research agencies.