MIT Model Lets AI "See" Crystal Structures in 3D to Find Better Materials
MIT researchers have built a multimodal AI model that pairs 3D atomic structures with natural language descriptions, allowing it to reason about crystalline materials and their properties. In a paper in Nature Communications, the team reports that it identified better material designs than models trained on text alone.
- MIT researchers built a multimodal AI model that combines 3D atomic structures with natural language descriptions of crystalline materials.
- According to the Nature Communications paper, it identified better material designs than models trained on text streams alone.
- In a solid-state lithium-ion battery electrolyte query, top results included materials independently identified as promising candidates.
- A model trained on AI-generated descriptions aligned geometric structures with text, while one trained on papers only did not.
- The researchers hope other materials scientists will use the approach to accelerate their work.
Why language models struggle with crystals
In principle, AI could make finding the right material for an application as simple as telling a large language model what is needed and receiving a designed crystal in return. According to MIT News, that approach does not work well yet, and the problem comes down to data.
Most materials consist of crystals, which are 3D periodic arrangements of atoms. Their structures are stored in specialized formats rather than the text streams that today's large language models are trained on. The experimental literature is also heavily biased toward a few popular classes of crystals, which makes AI outputs skew toward those materials.
Existing language models rely heavily on text from published papers and books. That limits their knowledge of the physical world, since many potential crystal structures sit in computational databases. Text-based models also lack an understanding of visual geometry. Crystalline materials have highly ordered atomic structures and make up more than 90 percent of solid materials on Earth, according to the article.
How the model works
The team created a vision-language model for crystalline materials. It combines 3D atomic structures with natural language descriptions, drawing on experimental research papers and first-principles computational databases. The resulting unified description covers the chemical composition and structure, as well as properties, processing considerations and potential applications.
The researchers trained a language model to create textual descriptions of crystal structures using contrastive learning. In this approach, models compare similar and dissimilar examples of structures and descriptions to reach a unified understanding. MIT compares this to aligning the visual and language regions of the brain.
The team then compared two versions of the model: one trained on a dataset containing the AI-generated descriptions, and one trained on text from papers only. The first learned to align geometric structures with text descriptions, while the other did not. "We found when we used that narrative dataset, the models are better at predicting the function," said Ju Li, the Carl Richard Soderberg Professor in Power Engineering in the Department of Materials Science and Engineering.
Results and battery electrolyte example
The researchers showed the model could take a natural language query describing a property and identify crystalline materials with that property, outperforming a baseline. In one test, they asked it to find materials suited for solid-state lithium-ion battery electrolytes. Its top results included materials that had been independently identified as promising candidates.
The paper, titled "Scaling multimodal materials representation learning with synthetic narratives," also reports that the model learned geometric representations of material structures in ways text-only models could not.
Li is joined on the paper by Yang Jeong Park, first author and former MIT postdoc; Mayank Kumaran, a recent University of Illinois Urbana-Champaign computer engineering graduate; Chia-Wei Hsu SM '26; and Elsa Olivetti, the Jerry McAfee (1940) Professor in Engineering in DMSE. The work was supported in part by the U.S. Defense Threat Reduction Agency through the University Research Alliance, the U.S. Defense Advanced Research Projects Agency, and the National Research Foundation of Korea.
What it could mean for materials science
Li compares the model to AlphaFold, an AI model that links protein structures with function and has helped advance biological research and drug development. "Structure is really at the heart of the materials science field," he said. "If you understand the structure, you can infer quite a lot about the properties and performance of materials."
Olivetti said the approach helps researchers see more deeply into three-dimensional structures and link structures to property and processing. "I hope this can be one of those step changes," she said.
The researchers believe the approach can work with many crystalline materials, and they hope other materials scientists will use it to accelerate their work. Li said: "I think this will be useful to anyone who practices materials science."
MIT News: Giving AI a “crystal clear” view of the materials world
