The crystal comparison publishes no score

The visible publication date on this MIT News piece is 18 October 2026. It gives no clock. The report says that asking a language model to hand back a suitable crystal from one request does not work well yet. The stated reason is data: most materials are crystals, three-dimensional periodic arrangements of atoms, stored in specialized formats rather than the text streams language models train on. The experimental literature is also biased toward a few popular classes, and the outputs skew with it. MIT researchers built a multimodal model that puts 3D atomic structures next to natural-language descriptions and generates text about those structures. A description covers composition and structure, and also properties, processing considerations, and possible applications.

The paper is in Nature Communications. Its title is Scaling multimodal materials representation learning with synthetic narratives. The report says that once the model can “see” atomic arrangements and bonding, it identifies better material designs than models trained on text streams alone. In one example, the researchers asked for materials suited to solid-state lithium-ion battery electrolytes. The top results included materials that had been independently identified as promising candidates. The report names none of those materials, does not say how many were ranked, and does not say by how much they beat text alone. It also says the model learned geometric representations that text models did not. That sentence has no count either.

Training uses contrastive learning: the model compares similar and dissimilar structures and texts until the two line up. They compared two versions. One was trained on a dataset that included the generated descriptions. The other was trained on text from papers only. The generated-description version learned to align geometric structures with text. The other did not. Ju Li says they synthesized narratives instead of relying only on existing papers because, with that narrative dataset, the models are better at predicting function. “Better” is his judgment. The report then says a natural-language query for a property can identify crystals with that property, outperforming a baseline. It does not say by how much, or what the baseline is.

Crystals make up more than 90 percent of solid materials on Earth. That sentence is about what solids are, not about this model’s accuracy. Li compares the model to AlphaFold, which he describes as linking protein structures with function and as having helped biological research and drug development. That is an analogy. This piece reports no biological experiment and no protein result. The authors include first author and former postdoc Yang Jeong Park, recent University of Illinois Urbana-Champaign graduate Mayank Kumaran, Chia-Wei Hsu, and Elsa Olivetti. Support came in part from the U.S. Defense Threat Reduction Agency through the University Research Alliance, from the U.S. Defense Advanced Research Projects Agency, and from the National Research Foundation of Korea. The funder list is not a list of delivered materials.