Teaching AI to See in 3D: How MIT's Crystal Model Outperforms Text-Only Systems at Materials Discovery

Introduction

In principle, finding the right material for an application could be as simple as ordering dinner. A scientist describes what they need — a solid that conducts lithium ions, survives thousands of charge cycles, and holds up under heat — and an artificial intelligence model serves back a perfectly designed crystal for the job. In practice, that vision has not worked well, and the reason is not that the models are too small. It is that most materials exist in a format today's large language models were never trained to read.

More than 90 percent of the solid materials on Earth are crystalline: three-dimensional periodic arrangements of atoms whose structures live in specialized database formats, not in the text streams that language models consume. The experimental literature that would otherwise teach such models is heavily biased toward a small set of popular, heavily studied material classes, so AI trained on that literature inherits the same narrowness. Now a team of researchers at MIT has built a new multimodal model that attacks both problems at once, teaching a system to read crystal structures the way a vision-language model reads an image — as geometry first, language second.

The result, published in Nature Communications, is a system that can identify better material designs than models trained on text alone, and that can answer natural-language questions about which crystals fit a given task. Coverage of artificial intelligence in science and engineering has tracked similar structure-aware models before; this one is aimed squarely at the periodic tables of the material world.

The Problem: Why Text-Only Models Fail at Materials

Materials scientists have spent the last decade hunting for new crystals that could become the products of the future. Over that period, AI has reshaped the search: machine-learning models now accelerate virtual screening, property prediction, generative design, and active learning. Most of those models treat crystalline solids as three-dimensional graphs, letting neural networks capture structural features without manual engineering.

The arrival of large language models offered another tool, and with it a question researchers call the fifth research paradigm in materials science — one in which AI systems integrate human knowledge, symbolic reasoning, and data-driven patterns. But that promise has been constrained by a limitation that has nothing to do with model size: text-only LLMs have no genuine grasp of visual geometry. They can describe a crystal in words, but they cannot see one.

According to the MIT team, two concrete obstacles stand in the way. The first is data. A vision-language model for images might be trained on hundreds of millions of image-text pairs, and molecular researchers can draw on billions of molecular structures. Crystal databases, by contrast, hold only millions of entries. The second obstacle is bias. Crystal data reflects a sociological tendency to concentrate research on a small number of so-called hot materials, and that concentration shows up in the language: papers describe those popular materials in rich detail, while thousands of theoretically interesting structures on the thermodynamic convex hull receive almost no description at all. A model trained on that literature inherits the bias without knowing it.

The Solution: Teaching a Model to See Crystals in Three Dimensions

The MIT researchers — led by Ju Li, the Carl Richard Soderberg Professor in Power Engineering, with first author Yang Jeong Park, a former MIT postdoc — built what they call a contrastive language-crystals model, or CLaC. It is a multimodal contrastive-learning framework that aligns three-dimensional atomic structures with natural language descriptions, in much the same way that CLIP aligns images with captions.

The architecture has two branches: a crystal-graph encoder that encodes a structure as a geometric graph that preserves spatial invariance, and a text encoder. Both are trained into a shared latent space so that a crystal structure and a written description of it end up close together. The team used an approach called contrastive learning, in which the model compares similar and dissimilar examples of structures and descriptions until the two modalities click into a unified understanding — the researchers liken it to aligning the visual and language regions of the brain.

The training data is where the innovation is sharpest. Rather than scraping only what papers happen to contain, the team generated synthetic narratives: language models produced rich textual descriptions of crystal structures drawn from first-principles computed properties and public databases such as the Materials Project. Each description captures not just chemical composition and structure, but properties, processing considerations, and potential applications. Those narratives cover roughly 300 times more diverse crystals than the human-written academic literature does, which removes the concentration bias that limits text-only training.

After training, the team compared two versions of the model — one trained on the synthetic narratives, one trained on raw paper text. Only the version trained on synthetic descriptions successfully learned to align geometric structures with text. The paper reports that models trained on physically grounded synthetic narratives learn a structured association between language and geometry, while models trained on literature text do not.

That distinction produced a testable result. When researchers asked the model to identify materials suited for solid-state lithium-ion battery electrolytes, its top results included materials that had already been independently identified as promising candidates by other methods. The model was also evaluated on its capacity for zero-shot inverse design: given a natural-language query describing a desired property, it could retrieve crystal structures that matched, outperforming a baseline.

The Broader Picture: From Protein Folds to Battery Electrolytes

Ju Li draws an explicit comparison to AlphaFold, the AI model that linked protein structure to function and transformed biological research and drug development. The analogy is apt: AlphaFold succeeded because it treated structure as the central object, and CLaC is making the same bet for crystalline solids. "Structure is really at the heart of the materials science field," Li said in the MIT announcement. "If you understand the structure, you can infer quite a lot about the properties and performance of materials."

The practical stakes are concrete. Solid-state electrolytes are one of the most-watched problems in battery research, because a successful one could unlock cells that are safer and more energy-dense than today's lithium-ion designs. A model that can screen a far larger combinatorial space than human researchers can — and that can be queried in plain language rather than specialized database syntax — compresses the early stage of that search.

The work was supported in part by the U.S. Defense Threat Reduction Agency and the Defense Advanced Research Projects Agency, alongside the National Research Foundation of Korea, which points to the national-security dimension of materials discovery as well as the commercial one. The researchers say the approach is not limited to one class of crystals and hope other materials scientists will adopt it. Whether CLaC becomes the field's AlphaFold moment is still an open question — the model has to prove itself across many more material families before that claim holds. What the paper establishes is more modest and more durable: that for the first time, a model can read a crystal the way it reads a sentence, and answer a question about what it is for.

Images

References

• Zach Winn, MIT News, "Giving AI a crystal clear view of the materials world," October 18, 2026 — https://news.mit.edu/2026/giving-ai-crystal-clear-view-of-materials-world-1018
• Yang Jeong Park, Mayank Kumaran, Chia-Wei Hsu, Elsa Olivetti, Ju Li, "Scaling multimodal materials representation learning with synthetic narratives," Nature Communications 17, 10247 (2026) — https://doi.org/10.1038/s41467-026-76378-7