MIT researchers reported a multimodal AI model in a paper published in Nature Communications that links 3D crystal structures with natural-language descriptions to identify materials for specific uses. The model combines atomic structures with descriptions drawn from experimental research papers and first-principles computational databases. Its unified descriptions include chemical composition, structure, properties, processing considerations, and potential applications. The team said this approach addresses limitations in text-only models, which have difficulty using crystal structures stored in specialized formats and may reflect research literature biased toward a few popular crystal classes.

The researchers tested the model on a request for materials suited to solid-state lithium-ion battery electrolytes. Its top results included materials that had been independently identified as promising candidates, and it outperformed a baseline when matching natural-language property queries to crystals. The researchers also found it learned geometric representations of material structures that text-only models did not.

To build the system, the team trained a language model to describe crystal structures using contrastive learning, a method that aligns related examples by comparing similar and dissimilar pairs. They then compared a version trained on a dataset of AI-generated descriptions with one trained only on text from papers. The description-trained model aligned geometric structures with text, while the paper-text-only model did not. Ju Li, a professor of materials science and engineering at MIT, said the team synthesized narratives to help models work broadly across materials and found they improved function prediction.

The researchers said crystals have highly ordered atomic structures and make up more than 90 percent of solid materials on Earth. Existing systems often represent crystals as 3D graphs, while language models trained on papers and books have limited access to structures in computational databases and lack visual geometry understanding. Li compared the approach with AlphaFold, which links protein structures with function. The team hopes materials scientists will use the method across many crystalline materials. The research received partial support from the U.S. Defense Threat Reduction Agency through the University Research Alliance, the Defense Advanced Research Projects Agency, and the National Research Foundation of Korea.

MIT researchers reported a multimodal AI model in a paper published in Nature Communications that links 3D crystal structures with natural-language descriptions to identify materials for specific uses. The model combines atomic structures with descriptions drawn from experimental research papers and first-principles computational databases. Its unified descriptions include chemical composition, structure, properties, processing considerations, and potential applications. The team said this approach addresses limitations in text-only models, which have difficulty using crystal structures stored in specialized formats and may reflect research literature biased toward a few popular crystal classes.

The researchers tested the model on a request for materials suited to solid-state lithium-ion battery electrolytes. Its top results included materials that had been independently identified as promising candidates, and it outperformed a baseline when matching natural-language property queries to crystals. The researchers also found it learned geometric representations of material structures that text-only models did not.

To build the system, the team trained a language model to describe crystal structures using contrastive learning, a method that aligns related examples by comparing similar and dissimilar pairs. They then compared a version trained on a dataset of AI-generated descriptions with one trained only on text from papers. The description-trained model aligned geometric structures with text, while the paper-text-only model did not. Ju Li, a professor of materials science and engineering at MIT, said the team synthesized narratives to help models work broadly across materials and found they improved function prediction.

The researchers said crystals have highly ordered atomic structures and make up more than 90 percent of solid materials on Earth. Existing systems often represent crystals as 3D graphs, while language models trained on papers and books have limited access to structures in computational databases and lack visual geometry understanding. Li compared the approach with AlphaFold, which links protein structures with function. The team hopes materials scientists will use the method across many crystalline materials. The research received partial support from the U.S. Defense Threat Reduction Agency through the University Research Alliance, the Defense Advanced Research Projects Agency, and the National Research Foundation of Korea.