Researchers at MIT have built an artificial intelligence model that can read the 3D structure of crystals. The goal is to help scientists find the right material for a specific job.
The work is described in a new paper in Nature Communications. MIT News reported on the research on October 18, 2026.
Today’s large language models learn mostly from text. But most solid materials are 3D arrangements of crystals, which do not fit neatly into text-based training.
Published research also focuses heavily on a few popular types of crystals. That bias pushes AI answers toward those same materials.
Crystals make up more than 90 percent of solid materials on Earth. Over the last decade, AI has been used to propose and design many new crystal structures.
How the Model Works
The MIT model combines 3D atomic structures with plain-language descriptions. Those descriptions draw on experimental research papers and computational databases.
Each description covers a material’s chemical makeup and structure. It also covers its properties, how it is processed and where it might be used.
The team used a method called contrastive learning. The model compares similar and different pairs of structures and descriptions to build a shared understanding of both.
The researchers then tested two versions of the model. One was trained on AI-generated descriptions, and the other was trained only on text from papers.
The version trained on AI-generated descriptions learned to match geometric structures with text. The text-only version did not.
“We wanted to create an AI model that works broadly across materials,” said Ju Li, an MIT professor of materials science and nuclear science. He said the models trained on the narrative dataset were better at predicting function.
Battery Material Test Results
The researchers gave the model plain-language requests describing a desired property. The model then identified crystalline materials with that property and outperformed a baseline model.
In one test, the team asked for materials suited to solid-state lithium-ion battery electrolytes. Its top results included materials that had already been independently identified as promising.
Li compared the new model to AlphaFold, an AI system that links protein structures to their function. AlphaFold has helped advance biology research and drug development.
The paper’s first author is Yang Jeong Park, a former MIT postdoc. Other authors include MIT professor Elsa Olivetti, Chia-Wei Hsu of MIT and Mayank Kumaran, a recent University of Illinois Urbana-Champaign graduate.
Funding came in part from the U.S. Defense Threat Reduction Agency, the U.S. Defense Advanced Research Projects Agency and the National Research Foundation of Korea.
The researchers say the approach can work with many types of crystalline materials. They hope other materials scientists will use it in their own work.
“Being able to see a three-dimensional crystal and how its atoms are connected is super foundational,” Li said. “I think this will be useful to anyone who practices materials science.”
