MIT model pairs 3D crystal structures with text to help AI reason about materials

Researchers at MIT, working with collaborators, have built a multimodal AI model that connects three-dimensional crystal structures with natural-language descriptions so that AI systems can reason about crystalline materials and their properties. The study, from the group of Ju Li in MIT's Department of Materials Science and Engineering, appears in the peer-reviewed journal Nature Communications. The paper names the system CLaC, for Contrastive Language-Crystal model.

According to MIT News, current large language models struggle with materials because crystal structures are kept in specialized formats rather than the text these models learn from, and because the experimental literature focuses on a small number of popular crystal classes. The paper notes that crystal databases hold only millions of entries, far fewer than the data behind vision-language models. To compensate, the team paired structures from the Materials Project with GPT-synthesized narratives grounded in computed properties, alongside paragraphs mined from papers. The authors report that the synthetic narratives cover a set of crystals 300 times more diverse than the human-written literature. A graph neural network encodes the structures, a text encoder handles the language, and contrastive learning places both in a shared representation space.

In zero-shot tests, the version trained on synthetic narratives matched structures to the right composition description with 95.4 ± 0.2% accuracy, compared with 13.5 ± 0.3% for an untrained model and 17.8 ± 0.3% for a baseline built on Gemma-3-1b-it, the paper says. A version trained only on text extracted from papers performed largely like random selection when asked to retrieve materials matching a target property. In a battery example, the researchers searched for lithium-ion solid-state electrolyte candidates using a high band gap criterion alone; all 30 top-ranked materials contained lithium, with an average band gap of 3.23 eV. Li6PClO5, ranked fourth, had been flagged as promising in independent work, and Li4SiO4, ranked ninth, had earlier been made as a thin-film solid electrolyte.

The authors also describe clear limitations. Accuracy on classifying crystal systems was only 34.0 ± 0.4%, roughly on par with the Gemma baseline, which they link to a small training batch size of 16 and a heavily imbalanced distribution of crystal systems. Ionic conductivity, a key property for electrolytes, is absent from the synthetic dataset, so it could not be used in the battery query. The paper adds that the field lacks standard multimodal benchmarks, making direct comparisons between models difficult, that instruction-based materials discovery with the model remains future work, and that more research is needed on how dataset composition and text generation shape retrieval behavior and possible biases.

"Structure is really at the heart of the materials science field", Li said in the MIT News release, adding that understanding structure allows researchers to infer much about a material's properties. He compared the model to AlphaFold, which relates protein structures to function. Co-author Elsa Olivetti said the approach links structure to property and processing, and added: "I hope this can be one of those step changes". The other authors are first author Yang Jeong Park, a former MIT postdoc, Mayank Kumaran, and Chia-Wei Hsu. According to MIT News, the work was partly funded by the National Research Foundation of Korea and by two US defense agencies, the Defense Threat Reduction Agency and the Defense Advanced Research Projects Agency.