For a language model, describing a crystal is by no means natural. Large generative systems have primarily been trained on text, images, and code; crystalline matter, on the other hand, is defined by three-dimensional, periodic arrangements of atoms, distances, bonds, and symmetries that are typically stored in specialized formats. This is a significant gap: in materials science, two substances with similar compositions can behave very differently precisely because of their structure.
A team at the Massachusetts Institute of Technology has developed a multimodal model designed to bridge this divide. The system combines the 3D representation of atomic structures with natural language descriptions drawn from experimental literature and computational databases based on first-principles calculations. The goal is not simply to pair the name of a compound with a property, but to provide the model with the foundation to link composition, lattice geometry, expected behavior, processing conditions, and potential applications.
The findings were described in a study published in Nature Communications. According to the researchers, integrating atomic structure makes it possible to propose better material designs compared to models trained solely on text streams. In a test focused on electrolytes for solid-state lithium-ion batteries, substances that had already independently emerged as promising were among the top candidates ranked by the model.
Why text data is not enough
The idea of entrusting materials discovery to AI has been discussed for some time: specify a requirement—such as stability, conductivity, or compatibility with a given manufacturing process—and the system is expected to narrow down an enormous search space to a viable shortlist for testing. In practice, however, the ability to read scientific papers alone does not equate to the ability to understand a material.
Structure is at the heart of the discipline because it holds decisive clues about a crystal's properties and performance. The arrangement of atoms and their bonding patterns are not illustrative details to be tacked onto a description: they are pieces of information that help determine how the material transports ions or electrons, how it responds to stress, and which phases it can adopt. Ju Li, a professor in the Department of Materials Science and Engineering at MIT and one of the study's authors, summarizes this step as the need to create a connection for the model between atomic structure, properties, and manufacturing processes.
Another limitation involves the quality and distribution of available information. Experimental literature tends to concentrate on a small number of heavily studied crystal classes. If a model learns primarily from that corpus, it can return suggestions skewed toward the most represented material families, rather than effectively exploring less-documented alternatives. Adding structural data provides a signal that goes beyond the mere recurrence of words in papers and databases.
The work from MIT therefore aims to transform the crystal from an object described verbally into one that the model can also consider in its three-dimensional form. The researchers report that the system has learned geometric representations of structures that a purely text-based approach could not achieve. This provides the foundation that lends greater credibility to shifting from literature review to genuine assisted screening or design.
The case of solid-state batteries
The demonstration on electrolytes for solid-state lithium-ion batteries is significant because this is a field where material selection is a central constraint. The electrolyte must satisfy diverse, and sometimes competing, requirements that cannot be properly inferred from an isolated chemical formula. The test does not prove that AI has single-handedly discovered a new production-ready material, nor does it replace laboratory verification. It does show, however, that by also looking at atomic organization, the model succeeded in surfacing candidates already identified through independent research paths.
It is a more cautious and useful form of validation than promises of full automation. Such a system can become a tool for setting priorities, formulating hypotheses, and narrowing down the number of avenues to explore experimentally. Confirming stability, synthesizability, costs, process yields, and real-world performance remains the work of chemists and engineers. Furthermore, relevant properties do not always depend solely on the ideal crystal: defects, impurities, interfaces, and manufacturing procedures can profoundly alter the final outcome.
This is precisely why including processing details in the model is a significant step. Materials discovery is not just about finding an elegant structure on paper: a candidate can be theoretically compelling while remaining difficult to produce, unstable, or unsuitable for device integration. Linking structural data with information gathered from experimental literature can help AI balance these dimensions without reducing the evaluation to a single predicted value.
A model for navigating complexity, not an autonomous lab
Alongside Li, the study was co-authored by Yang Jeong Park, lead author and former MIT postdoc, Mayank Kumaran, Chia-Wei Hsu, and Elsa Olivetti. Their contribution is part of a rapidly growing field that seeks to apply generative and multimodal models to chemistry and materials science. The difference, in this case, lies in the focus on a type of information that has historically been peripheral to language models: the three-dimensional geometry of crystals.
The practical impact will depend on how well such approaches can generalize beyond known data. Recognizing candidates validated by other studies is an encouraging sign, but the most demanding test will be generating novel, testable, and genuinely useful hypotheses for specific industrial or scientific problems. Then there is the issue of access and integration: to make a meaningful impact on research timelines, these tools will need to interface with reliable databases, simulations, and experimental workflows, rather than merely producing a convincing list.
MIT's work nevertheless points in a concrete direction. If AI is to help design matter, it must be able to treat structure as primary information, alongside the words scientists use to describe it. For crystalline materials, the ability to “see” atoms in space can make the difference between an engine that summarizes what is known and one that helps choose where to look next.
