Researchers at a US technical university have built a multimodal AI model that pairs the three-dimensional atomic structures of crystalline materials with natural-language descriptions. Crystals make up more than 90 per cent of solid materials on Earth, but most language models trained on text have little knowledge of their 3D geometry. The aim is to help AI link a material's structure to its properties, processing and potential uses.
The team used contrastive learning, which compares matching and non-matching structure-and-text pairs so that the model builds one shared understanding of both. Descriptions were generated from experimental research papers and from first-principles computational databases, and each covers composition, structure, properties, processing considerations and potential applications. The researchers compared a model trained on this AI-generated narrative dataset with one trained only on text from papers. Dataset sizes were not given.
The narrative-trained model learned to align geometric structures with text descriptions, and the paper-text-only model did not. For solid-state lithium-ion battery electrolytes, its top results included materials that other researchers had independently identified as promising. Given a plain-language query describing a property, it found crystalline materials with that property and outperformed a baseline that was not named.
Ju Li, a professor of power engineering and the senior researcher, said structure is at the heart of materials science and that models trained on the narrative dataset were better at predicting function. Elsa Olivetti, a professor of engineering and co-author, said she hopes it can be one of those step changes. The first author is Yang Jeong Park, a former postdoctoral researcher. The work was published in a peer-reviewed journal.
The limits are plain. No quantitative metrics are reported, so the size of the gains cannot be judged. The researchers believe the approach can work across many crystalline materials, but that is not demonstrated. How well the model generalises beyond the cases tested, and how accurate the AI-generated descriptions are, are not discussed.
