An MIT AI model that "sees" crystal structures picks better materials than text-only models

Researchers at MIT have built an AI model that combines the 3D atomic structure of crystals with written descriptions of their properties, and it picked out better materials than models trained on text alone, MIT News reported on 18 October. The work is published in Nature Communications.

Language models cannot see a crystal

Most solid materials are crystals, regular 3D arrangements of atoms. Their structures are stored in special data formats, not in the text that large language models learn from, and the research literature is heavily skewed toward a few popular crystal families. As a result, asking a chatbot for the right material for a job tends to return the usual suspects.

The MIT team trained a vision-language model for crystals using contrastive learning, in which the model compares matching and mismatching pairs of structures and descriptions. Because published papers cover so few materials, the researchers generated their own descriptions from computational databases, covering composition, structure, properties, processing and possible uses.

“If you understand the structure, you can infer quite a lot about the properties and performance of materials.” — Ju Li, Carl Richard Soderberg Professor in Power Engineering, MIT Department of Materials Science and Engineering

Battery electrolytes as a test

A version trained on the generated descriptions learned to link geometry with text; a version trained only on text from papers did not. When the team asked the model in plain language for materials suited to solid-state lithium-ion battery electrolytes, its top results included materials that other researchers had independently identified as promising. Li compares the approach to AlphaFold, which links protein structure to function.

The group laid the groundwork in an earlier preprint describing CLaC, a model trained on about 126,000 synthetic crystal-text pairs. Better AI for crystals fits a wider push in materials science to find materials faster than trial and error allows.

A tool, not yet a discovery

The study shows the model can rank known materials sensibly; it has not yet produced a new material that was then made and tested. “I think this will be useful to anyone who practices materials science,” Li said. The team hopes other researchers will adopt it. The work was supported in part by the US Defense Threat Reduction Agency, DARPA and the National Research Foundation of Korea.