Finding a material for a particular job requires more than matching words in scientific papers. A crystal’s three-dimensional arrangement of atoms can help determine what the material does, but that structure is often recorded in formats that ordinary language models cannot readily interpret. MIT researchers have built a model designed to connect those atomic arrangements with descriptions of materials and their potential uses.
Why text alone leaves gaps
Crystals are highly ordered atomic structures, and they make up more than 90 percent of solid materials on Earth. Materials scientists use AI to propose and design new crystalline structures, but many existing approaches represent them as 3D graphs rather than as the text streams used to train large language models.
Language models can draw on published papers and books, yet many possible crystal structures are held in computational databases. The experimental literature also focuses heavily on a few popular classes of crystals. That imbalance can steer a text-trained model toward familiar materials and leave it with limited understanding of the geometry that distinguishes one structure from another.
For materials research, that missing connection matters. A description of chemical composition alone may not convey how atoms are arranged or bonded. Those geometric features can help explain a material’s properties and performance.
Giving the model a shared language for structure and text
The MIT team created a vision-language model for crystalline materials. It combines 3D atomic structures with natural language descriptions drawn from experimental research papers and first-principles computational databases. Its descriptions bring together information about composition and structure, along with properties, processing considerations and possible applications.
To teach the model to connect structures and words, the researchers used contrastive learning. The system compares examples of structures and textual descriptions, learning which examples go together and which do not. The aim is a shared representation that lets the model relate a crystal’s geometry to what people say about its behavior and uses.
The team also generated textual narratives of crystal structures. This was meant to broaden the model’s descriptions beyond the material classes that dominate experimental papers. As Ju Li explained, the researchers wanted a model that worked broadly across materials, so they synthesized narratives instead of relying only on existing papers.
What the researchers found
The researchers compared a version trained on AI-generated descriptions with one trained on text from papers alone. The narrative-trained model learned to align geometric structures with text descriptions; the text-only version did not. The team also found that the narrative dataset helped models predict material function better.
In another test, the researchers gave the model a natural-language request for a property and asked it to identify crystalline materials that fit. The model outperformed a baseline. When asked for materials suited to solid-state lithium-ion battery electrolytes, its highest-ranked results included candidates that had been identified independently as promising.
These findings suggest that including structural information can improve how AI searches for materials with desired properties. They do not mean that a language model can replace experimental research. Instead, the model offers a way to connect information across structures and descriptions, helping researchers narrow down candidates to investigate.
A possible tool for materials research
Materials scientist Elsa Olivetti described the approach as a way to see more deeply into three-dimensional structures and link them to properties and processing. The team hopes the method will be useful across many crystalline materials and that other materials scientists will use it to accelerate their work.
Li compared the model’s role to AlphaFold, which links protein structures with function and has helped advance biological research and drug development. The comparison points to a similar goal: making structural information easier to connect with function. For materials scientists, the practical promise is a model that can consider both what a crystal looks like at the atomic level and what researchers want it to do.
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. The study was published in Nature Communications.