MIT builds AI that can see crystal structures in 3D
A new model combines atomic geometry with language so machines can help discover better materials.

An AI model that reads 3D atomic structures and text together to help scientists discover better materials faster
- Feed it a 3D crystal layout and its text description
- Compare similar and different examples together
- Learn how geometry connects to language descriptions
- Suggest crystals that match what you need
Seeing crystals in three dimensions
Materials scientists spend years hunting for the right crystals — highly ordered arrangements of atoms that make up more than 90 percent of solid materials on Earth — but a new AI model from MIT researchers could speed that search dramatically by letting machines "see" atomic structures the way a human would.
Published in Nature Communications, the system takes 3D atomic layouts and pairs them with natural-language descriptions pulled from experimental papers and computational databases, creating a richer understanding than models trained on text alone 1.
Most existing AI approaches treat crystals as 3D graphs, but they miss the visual and geometric nuance that matters when you want to predict how a material will behave 1. The MIT team wanted to close that gap.
How the model learns
The researchers trained the system using a technique called contrastive learning, in which the model compares similar and different examples of crystal structures alongside their textual descriptions, gradually learning to align the two until it understands how geometry maps to language — much like how the visual and language regions of the brain work together 1.
Instead of relying solely on what had already been published, the team synthesized new narratives that wove together composition, structure, properties and potential uses into unified descriptions 1. That choice mattered: when tested against a version trained only on paper text, the narrative-trained model was better at predicting function .
What the new AI sees that older models miss
Based on MIT research published in Nature Communications; older AI described by authors as baseline text-only approach.
What it can do
When the researchers asked the model to find crystals suited for solid-state lithium-ion battery electrolytes, its top results lined up with materials that other researchers had independently flagged as promising 1. The model also learned geometric representations of crystal structures that text-only systems could not, pointing to a deeper grasp of atomic arrangement and bonding patterns .
Ju Li, the Carl Richard Soderberg Professor in Power Engineering at MIT, said the work was driven by the belief that structure lies at the heart of materials science 1. "If you understand the structure, you can infer quite a lot about the properties and performance of materials," he said .
The research team also included first author Yang Jeong Park, a former MIT postdoctoral researcher; Mayank Kumaran, a recent computer engineering graduate from the University of Illinois Urbana-Champaign; Chia-Wei Hsu, who holds a SM from MIT; and Elsa Olivetti, the Jerry McAfee Professor in Engineering at MIT's Department of Materials Science and Engineering.
Why it matters
Finding new materials has traditionally been a slow, trial-heavy process. By giving AI a clearer picture of how atoms are arranged in 3D and linking that to real-world properties and applications, the model could help researchers identify better candidates faster — potentially accelerating development of batteries, semiconductors and other technologies that depend on advanced crystals 1.
Olivetti said she hopes the approach will become a useful tool the broader community adopts as the field moves toward deeper algorithmic integration 1. "There are sometimes these step changes within materials science around algorithmic integration toward useful tools, which the community uses to further advance the field," she said .
The road ahead
For now, the system has been tested on specific property searches and battery-related materials, but the researchers say the approach is designed to work broadly across classes of crystalline compounds rather than targeting just one application 1. As materials datasets grow and multimodal models mature, the same framework could eventually let scientists describe what they need in plain language and receive a shortlist of crystal structures that fit .
That vision — a future where finding the right material is as simple as asking the right question — is still ahead, but this model marks a meaningful step toward making it real.
Key facts
- Crystals make up more than 90 percent of solid materials on Earth.
- The model was trained using contrastive learning to align 3D atomic structures with text descriptions.
- It outperformed text-only models when asked to find materials for solid-state lithium-ion battery electrolytes.
- The research was published in Nature Communications.
- MIT researchers synthesized new narrative descriptions rather than relying solely on existing paper text.
Sources
If this story concerns you or your company, or uses something that belongs to you, ask for a correction or for it to be taken down. How requests work
Keep reading
How it worksNew method makes AI text detectors 93% accurate, catching disguised AI writingHerald of AI
New method makes AI text detectors 93% accurate, catching disguised AI writing
A training trick lets detection tools adapt in real time so they stay accurate as AI and human writing styles keep changing.
How it worksMIT researcher wants AI to explain itself instead of hiding how it worksHerald of AI
MIT researcher wants AI to explain itself instead of hiding how it works
Jacob Andreas is training AI to talk about what it learned — giving everyday users a way to understand how these systems make decisions, not just relying on complex debugging tools.
Apps and productsMIT's AI copilot lets engineers guide structural designs in real timeHerald of AI
MIT's AI copilot lets engineers guide structural designs in real time
A new tool shows building-designers the draft AI is making and asks their input — producing stronger, buildable structures in just 15 extra seconds.