MIT researchers train an AI model to link crystal structure with materials function

MIT researchers have developed a vision-language model designed to help identify crystalline materials for particular applications. It pairs three-dimensional atomic structures with natural-language descriptions of materials, aiming to connect a crystal’s geometry with its properties, processing and possible uses.

The approach addresses a mismatch between how materials are represented and how language models learn. Many solids have highly ordered crystal structures, but conventional language models are trained primarily on text. The research team also points to an imbalance in the experimental literature, which concentrates on a limited set of popular crystal classes and can skew what text-based systems recommend.

To build the model, the researchers used contrastive learning to align crystal structures with textual descriptions. They generated narratives drawing on experimental research papers and first-principles computational databases, then compared a model trained on those descriptions with one trained on paper text alone. MIT reports that the narrative-trained system learned to align geometric structure with language, while the text-only version did not.

In tests, the multimodal model outperformed the text-only baseline at identifying materials with requested properties. When prompted to find candidates for solid-state lithium-ion battery electrolytes, its top results included materials already identified independently as promising candidates, according to the researchers. The result is evidence of improved retrieval against that baseline—not a report that the system discovered or experimentally validated a new electrolyte.

The work, published in Nature Communications under the title “Scaling multimodal materials representation learning with synthetic narratives,” was led by first author Yang Jeong Park, a former MIT postdoctoral researcher, with MIT professor Ju Li and other collaborators. The research received support from the U.S. Defense Threat Reduction Agency, the Defense Advanced Research Projects Agency and the National Research Foundation of Korea.

Li likened the model’s ambition to AlphaFold’s linking of protein structure and function. The MIT team says the method could be useful across crystalline materials, but the reported results establish a benchmark comparison and a candidate-identification task, not yet a general-purpose materials-discovery system tested in laboratory workflows.