In an ideal world, a materials scientist could describe the property they need, a battery electrolyte that conducts lithium well, say, and an AI model would suggest crystals that fit. Today's language models are poor at this, and an MIT team argues the reason is data. Materials are 3D arrangements of atoms, not text, and the published literature covers only a handful of popular crystal classes, so models trained on papers skew towards what has already been studied.
The team, led by Ju Li with first author Yang Jeong Park and colleagues including Elsa Olivetti, built a model that links crystal structures directly with language. Using contrastive learning, which teaches a model to pair matching examples and separate mismatched ones, it aligns 3D atomic structures with text descriptions, much as image models align pictures with captions. The key move was the training text. Rather than rely only on papers, the researchers generated descriptions for structures from both the literature and first-principles computational databases, covering composition, structure, properties, processing and possible uses. A version trained on these synthetic descriptions learned to align geometry with text; one trained on paper text alone did not.
The payoff is search by description. Given a plain-language query about a property, the model returned crystalline materials with that property and beat a baseline. Asked for solid-state lithium-ion battery electrolytes, its top results included materials others had already identified as promising. Li compares the ambition to AlphaFold, which linked protein structure to function, and the paper is published in Nature Communications.
For researchers, the idea matters because crystals make up more than 90% of solid materials on Earth, and computational databases already hold far more candidate structures than anyone has written about. A tool that can read those structures and answer questions in plain language would make that store of data easier to use, alongside graph-based models that already propose new crystals.
The evidence so far is thin for practical use. The MIT article gives no performance figures, does not name the baseline, and does not say how many candidates were checked. Recovering materials that experts already knew about shows the search works; it does not show the model can find new ones, which is the test that matters. The training descriptions are themselves AI-generated, so errors in them could pass into the model. Independent tests on materials outside the training data will decide whether this is the step change its authors hope for.