In principle, artificial intelligence could make finding the right material for a job as easy as ordering dinner: describe what you need to a large language model, and it hands you a crystal designed for the task. In practice, that scenario has remained out of reach, because the models behind today's chatbots have never actually seen what most materials look like.
MIT researchers report they have taken a concrete step toward fixing that. In a new paper published in Nature Communications, a team from MIT's Department of Materials Science and Engineering (DMSE) describes a multimodal AI model that combines three-dimensional atomic structures with natural-language descriptions, letting the system reason about crystal geometry and text at the same time. According to MIT News, the model identifies better material designs than systems trained on text alone by allowing the AI to see the atomic arrangements and bonding patterns that determine how materials behave. The work adds to a busy stretch of AI industry coverage focused on models that move beyond text and into the physical world.
Why Text-Trained Models Stumble on Matter
The obstacle is fundamentally about data. Most materials are crystalline: three-dimensional, periodic arrangements of atoms whose structures are stored in specialized computational formats rather than the text streams that large language models are trained on. On top of that, the experimental literature is heavily biased toward a few popular classes of crystals, which skews the outputs of any AI trained primarily on published papers.
The result is a blind spot with real consequences. More than 90 percent of solid materials on Earth are crystalline, and many of the structures that could matter for next-generation batteries, electronics, or catalysts live in computational databases that language-only systems effectively cannot read.
Teaching a Model to Align Structure With Language
The MIT team's answer was to build what is effectively a vision-language model for crystals. The researchers trained a language model to produce textual descriptions of crystal structures using contrastive learning, an approach in which a model compares similar and dissimilar examples of structures and descriptions until it develops a unified understanding — a process MIT compares to how the brain aligns its visual and language regions.
Crucially, the team chose to synthesize its own narratives about materials rather than relying solely on existing papers. Drawing on experimental research literature and first-principles computational databases, the model generated rich descriptions covering not just chemical composition and structure, but also properties, processing considerations, and potential applications.
"We wanted to create an AI model that works broadly across materials," said Ju Li, the Carl Richard Soderberg Professor in Power Engineering at MIT and a senior author on the paper. "That's why we decided to synthesize narratives rather than rely solely on existing papers. We found when we used that narrative dataset, the models are better at predicting the function."
The Test: Finding Solid-State Battery Electrolytes
To evaluate the approach, the researchers compared two versions of the model: one trained on the AI-generated structural descriptions and one trained only on text from research papers. The version trained on the synthesized descriptions successfully learned to align geometric structures with language, while the text-only version did not.
The team then asked the model to identify crystalline materials suited for solid-state lithium-ion battery electrolytes — a key challenge for next-generation batteries. According to MIT News, the model's top results included materials that had been independently identified as promising candidates, and the system outperformed a text-only baseline when given natural-language queries describing a property and asked to find crystals with that property.
"Structure is really at the heart of the materials science field," Li said. "If you understand the structure, you can infer quite a lot about the properties and performance of materials. We wanted to build the connection between a material's atomic structure and its properties and processing for the model."
Where This Fits in AI-Driven Materials Discovery
Over the past decade, AI has already reshaped how new crystals are proposed and designed; most of those systems, however, treat crystalline materials as three-dimensional graphs processed by specialized architectures, separate from the language models that power general-purpose AI. By teaching a language-based model to perceive geometry directly, the MIT work points toward assistants that could one day handle a materials request stated in plain English end to end.
The research team included first author Yang Jeong Park, a former MIT postdoc, along with Mayank Kumaran, a recent University of Illinois Urbana-Champaign computer engineering graduate, Chia-Wei Hsu, and Elsa Olivetti, the Jerry McAfee (1940) Professor in Engineering at MIT.
What Comes Next
The model remains a research system, and MIT has not announced any productization. But the implications are broad: solid-state electrolytes were one test case, and the same structure-plus-language approach could apply to catalysts, semiconductors, or structural materials. If the approach generalizes, the bottleneck in materials discovery may shift from generating candidate crystals to validating them in the lab — a shift experimentalists would likely welcome.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
