Searching for materials in plain English

Researchers at the Massachusetts Institute of Technology published a model this week, in Nature Communications, that lets you search for materials in plain language.

Ask for something with a given property and it returns candidate crystals, the ordered atomic structures that make up more than 90% of solid materials on Earth.

The method pairs each crystal's 3D atomic structure with a written description of its composition, properties and processing, and learns to line the two up. So you query the structure space with words instead of having to read structures yourself.

The honest limits: the paper reports no benchmark score, only that it beat a text-only baseline. The experimental literature leans heavily toward a few popular crystal families, which skews what the model surfaces. And the training descriptions are themselves machine-written, so the output is only as good as they are.

Still, the bottleneck it moves is real. Searching materials space used to require reading the structures, which meant you needed the training to read them. That gated the search to specialists. Put a plain-language query in front of it and the gate moves: the scarce skill is no longer reading the crystal, it's knowing which property to ask for.

Asked for solid-state battery electrolytes, its top picks included materials other groups had already flagged on their own. Suggestive, not proof. But the right kind of suggestive.

Read the primary source →