MIT Researchers Build AI Model That Reads Crystals in 3D to Find New Materials

Researchers at the Massachusetts Institute of Technology have built an artificial intelligence model that can “see” the three-dimensional structure of crystals and connect it to plain-language descriptions of what those materials can do. The work, described in a new paper in Nature Communications and announced by MIT News on Oct. 18, 2026, is aimed at a stubborn problem in materials science: getting AI to suggest the right material for a job, such as a better battery electrolyte.

The Problem: Language Models Can’t See Crystals

Most solid materials are crystals, meaning their atoms sit in repeating 3D patterns. According to MIT, crystalline structures make up more than 90% of solid materials on Earth. Those structures are usually stored in specialized data formats in computational databases, not in the kind of text that today’s large language models learn from.

That creates two gaps. First, a text-only model knows only what has been written down in papers and books, and much of what scientists know about possible crystals lives in databases instead. Second, the published literature is heavily weighted toward a few popular classes of crystals, so AI answers tend to drift toward those materials. Text models also lack any sense of geometry, which is central to how a material behaves.

“Structure is really at the heart of the materials science field,” Ju Li, MIT’s Carl Richard Soderberg Professor in Power Engineering in the Department of Materials Science and Engineering, said in the MIT announcement. “If you understand the structure, you can infer quite a lot about the properties and performance of materials.”

How the MIT AI Model Works

The team built what is known as a vision-language model for crystalline materials. Instead of relying only on existing papers, they used AI to generate detailed written descriptions, or “narratives,” for crystal structures drawn from both experimental research and first-principles computational databases. Each description covers the material’s chemical makeup and structure along with its properties, processing considerations and possible applications.

They then trained the model with contrastive learning, a technique in which the model compares matching and non-matching pairs of structures and descriptions until it learns which ones belong together. MIT compared the approach to aligning the visual and language areas of the brain.

To test the idea, the researchers trained two versions: one on the AI-generated narratives and one on text from papers alone. The version trained on the narratives learned to line up geometric structures with their descriptions; the paper-only version did not. “We found when we used that narrative dataset, the models are better at predicting the function,” Li said.

What It Found for Batteries

The researchers showed the model could take a natural-language request describing a property and return crystalline materials with that property, beating a baseline model. When asked for materials suited to solid-state lithium-ion battery electrolytes, its top results included materials that other researchers had already independently identified as promising candidates. Solid-state electrolytes are a major research target because they could make batteries safer and more energy-dense than today’s liquid designs.

Li compared the model to AlphaFold, the AI system that links protein structure to function and has accelerated biology and drug research. Elsa Olivetti, the Jerry McAfee (1940) Professor in Engineering, said such “step changes” happen in materials science when new algorithms become everyday tools. “This approach helps us see more deeply into three-dimensional structures and link structures to property and processing,” she said. “I hope this can be one of those step changes.”

Who Did the Work and Who Paid for It

The paper’s first author is Yang Jeong Park, a former MIT postdoc. Co-authors include Mayank Kumaran, a recent University of Illinois Urbana-Champaign computer engineering graduate; Chia-Wei Hsu, who earned a master’s degree at MIT in 2026; Olivetti; and Li. MIT says the work was supported in part by the U.S. Defense Threat Reduction Agency, the Defense Advanced Research Projects Agency (DARPA) and the National Research Foundation of Korea.

Why It Matters for Boston AI

The study is part of a wider push in Cambridge and Boston to use AI as a tool for science, not just for chatbots. This fall alone, the Simons Foundation and two AI labs committed $15 million to help mathematicians adapt to AI, and Microsoft unveiled a biology AI model. MIT and Harvard remain the anchors of the region’s research base; Harvard recently announced a $150 million research commitment.

The model is a research result, not a product. The MIT announcement did not say whether the code or data will be released publicly, and any material the model suggests would still need to be made and tested in a lab. For more AI research and Boston tech news, visit our AI news hub.

Featured photo: John Phelan / Wikimedia Commons, CC BY 3.0. Cropped and resized.

Sources

• MIT News: Giving AI a “crystal clear” view of the materials world (Oct. 18, 2026)
• Nature Communications
• MIT Department of Materials Science and Engineering
• Google DeepMind: AlphaFold