The Pulse
MIT’s CLaC Model Links Crystal Structures to Material Properties
MIT researchers introduce CLaC, an AI model that matches three-dimensional crystal structures with natural-language descriptions of their properties and uses. In tests, synthetic, property-grounded training descriptions helped the model ret
AI.info Team · Written with AI ·
From atomic geometry to a materials search
MIT researchers have built an AI model that connects a crystal’s three-dimensional atomic arrangement with descriptions of its properties and possible uses. Called Contrastive Language-Crystal, or CLaC, the system lets researchers search a database of crystalline materials with ordinary language rather than relying only on formulas or specialized structural files.
The study, published in Nature Communications on August 26, describes a way to pair crystal structures with text in a shared representation space. MIT News reported on the work on October 18. The researchers’ aim is to help a model associate how atoms are arranged with what a material can do, including how it might be processed and where it could be useful.
“By contrast, models trained solely on raw literature text fail to exhibit comparable semantic organization, underscoring the importance of synthetic data quality rather than scale alone.”
Yang Jeong Park, Mayank Kumaran, Chia-Wei Hsu, Elsa Olivetti, and Ju Li, authors of “Scaling multimodal materials representation learning with synthetic narratives”
Why the model uses synthetic descriptions
Crystals are usually represented as structured three-dimensional data, while language models learn from text. The research team’s approach trains a graph neural network to encode crystal structures and a language model to encode descriptions, then uses contrastive learning to bring matching structure-and-text pairs closer together in a shared space. The crystal encoder is PaiNN, and the text encoder is Galactica-125M.
The training set pairs roughly 126,000 materials from the Materials Project with descriptions generated by language models and grounded in calculated material properties. The researchers also tested a literature-based alternative, using text gathered from a corpus of four million scientific papers. That text produced about 400,000 sentence pairs, but those pairs represented roughly 400 inorganic compounds, illustrating how published research concentrates on a relatively narrow group of materials.
The comparison points to a practical obstacle: adding more scientific text does not necessarily teach a model the link between a specific crystal geometry and a useful property. CLaC’s synthetic descriptions explicitly connect structures with properties, processing details and applications, giving the model a more direct training signal.
A battery query tests the link
The researchers tested whether CLaC could retrieve candidate materials from a natural-language description of a desired property, without extra task-specific training. In a search related to solid-state lithium-ion battery electrolytes, the model returned lithium-containing materials with a high average band gap among its top 30 results. One candidate, Li₆PClO₅, ranked fourth and had been identified independently as promising.
That example has an important limit: the study’s synthetic descriptions did not include ionic conductivity, another property needed to assess an electrolyte. The model therefore searched using a high-band-gap condition, not a complete specification for a working battery material. The result shows that the model can retrieve candidates associated with a stated property; it does not show that CLaC has established a material’s performance in a battery.
The paper also reports that synthetic-text training outperformed a random retrieval baseline across the tested properties, while the literature-text version performed much closer to random selection. That distinction matters: the model’s gains come from how the text is paired with structures, not simply from giving it a larger pile of papers.
Single crystals set the current boundary
CLaC is designed for crystalline materials and does not currently handle polycrystals, defects or metal-organic frameworks, the authors write. Those cases matter because real materials can contain irregularities or multiple crystal grains that alter their behavior. The current system therefore covers only part of the materials researchers may need to evaluate.
The work presents CLaC as a search and representation tool, rather than a system that synthesizes new materials or independently verifies their properties. Its immediate contribution is a way to translate between atomic structure and the language scientists use to describe material behavior. For battery-related searches, the tested example ranked Li₆PClO₅ fourth among the retrieved candidates, while the query itself left ionic conductivity out.