Giving AI a “crystal clear” view of the materials world

MIT researchers have built a multimodal AI model designed to help scientists find crystalline materials suited to particular uses, MIT News reports. The work, published in Nature Communications, targets two weaknesses in large language models: they learn mostly from text, while materials are defined by 3D atomic arrangements, and the experimental literature leans heavily toward a few popular crystal classes.

The team paired 3D atomic structures with synthetic written narratives drawn from research papers and first-principles computational databases, covering composition, structure, properties, processing and possible applications. Using contrastive learning, which compares matching and mismatched examples of structures and descriptions, they trained the system to connect geometry with language. A version trained on these generated descriptions learned to align structures with text, while one trained only on text from papers did not.

Given plain-language property queries, the model beat a baseline at identifying suitable crystals. Asked about electrolytes for solid-state lithium-ion batteries, it returned top results that included materials independently flagged as promising.

Ju Li, an MIT professor of materials science and nuclear engineering, likens the tool to AlphaFold, which links protein structure to function, and expects it to be useful across the field. Co-author Elsa Olivetti hopes it could prove to be a "step change" for materials research. Other authors are first author Yang Jeong Park, a former MIT postdoc; Mayank Kumaran; and Chia-Wei Hsu. Funding came partly from the U.S. Defense Threat Reduction Agency, DARPA and the National Research Foundation of Korea.