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

MIT researchers have developed a multimodal AI model for crystalline materials that links 3D atomic structures to natural-language descriptions, and reported the work in a paper in Nature Communications. The team was led by Ju Li, the Carl Richard Soderberg Professor in Power Engineering in the departments of Materials Science and Engineering and Nuclear Science and Engineering. The first author is Yang Jeong Park, a former MIT postdoc, joined by 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 problem the work targets is a data mismatch. Most large language models learn from streams of text drawn from papers and books, but solid materials are mostly three-dimensional arrangements of crystals, which carry their information in geometry and bonding rather than prose. The experimental literature is also skewed toward a handful of popular crystal classes, so models trained on it tend to favor those materials in their outputs. Many candidate structures exist only in computational databases built from first-principles calculations, where no paper text describes them.
The MIT approach generates the missing text instead of waiting for it. The researchers used contrastive learning to train a language model to produce descriptions of crystal structures, comparing similar and dissimilar pairs of structures and descriptions until the two representations line up. The published account likens the mechanism to aligning the visual and language regions of the brain, which is the standard way vision-language models bind an image encoder to a text encoder. Each generated description covers chemical composition and structure along with properties, processing considerations and potential applications, giving the model a fuller picture of what makes a material useful.
The team then ran a controlled comparison: one model trained on a dataset containing the generated descriptions, another trained on paper text only. The version trained on generated descriptions learned to align geometric structures with text, while the text-only version did not, according to the researchers. They also reported that their model learned geometric representations of material structures in ways text-only models could not.
On a practical test, the researchers queried the model for materials suited to solid-state lithium-ion battery electrolytes. Its top results included materials that had been independently identified as promising candidates in separate work, and the model outperformed a baseline on identifying crystalline materials from a natural-language property query. The electrolyte case is a demanding one because a solid electrolyte must satisfy several requirements at once, including ionic conductivity, electrochemical stability against electrodes and mechanical behavior, so a model that surfaces credible candidates is doing more than keyword matching.
Li framed the result around structure as the organizing principle of the field, noting that understanding a material's structure allows one to infer a great deal about its properties and performance, and that the goal was to connect atomic structure to properties and processing. He compared the model to AlphaFold, the protein-structure prediction system that helped advance biological research and drug development, and said the aim was a model that works broadly across materials rather than on a narrow class. Olivetti described the work in terms of step changes in materials science, where an algorithmic tool gets integrated into community practice and helps researchers see more deeply into three-dimensional structures and tie them to property and processing.
Crystalline structures make up more than 90 percent of solid materials on Earth and have highly ordered atomic arrangements, which is why most prior AI work in this space represents them as 3D graphs. That representation is powerful for predicting a single property but has no natural interface to language. This new work suggests the gap can be bridged by generating the textual side of the pair rather than relying on literature alone. In practice, a model that can accept a plain-language property request and return candidate crystals is closer to a usable design tool for researchers who are not machine-learning specialists, and it opens the door to searching computational databases that no paper describes.
The researchers believe the approach can work with many crystalline materials and hope other materials scientists will adopt it to speed their work; Li described the ability to see a three-dimensional crystal and its atomic connectivity as foundational for anyone practicing materials science. 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.
A likely trade-off is that generated descriptions inherit whatever the underlying computational databases and models get right or wrong. If a simulation misestimates a property, the synthesized narrative can carry that error into training, and the contrastive alignment will still look clean. The reported evaluation also centers on a small number of tasks, a comparison against a text-only baseline, and an electrolyte query; broader validation across many property types and crystal classes would show how far the gains extend. The team is openly positioning the work as a tool for the community rather than a finished discovery engine.
Why it matters: Materials discovery teams that rely on text-based assistants have been limited by literature bias and by the lack of geometric understanding in language models. A model that pairs structure and language, as reported here, points toward assistants that can take a plain-language property request and return plausible crystal candidates, including ones documented only in computational databases. If the approach holds across material classes, inference suggests the near-term effect is less about automating discovery end to end and more about compressing the early screening stage for researchers working on batteries and other solid-state technologies.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
New research can help AI models understand crystalline materials in 3D and reason about their properties.