The paper's authors, alongside Li, are Yang Jeong Park (first author and former MIT postdoc), Mayank Kumaran (a recent University of Illinois Urbana-Champaign computer engineering graduate), Chia-Wei Hsu SM '26, and Elsa Olivetti, the Jerry McAfee (1940) Professor in Engineering.

How did they teach a language model geometry?

The team trained a language model to write textual descriptions of crystal structures using contrastive learning — a technique where the model compares similar and dissimilar examples of structures and their descriptions until it forms one unified understanding. Li likens this to aligning the visual and language regions of the brain.

The payoff showed up in a head-to-head test. The researchers compared two versions of the model: one trained on AI-generated narrative descriptions, the other trained only on text from published papers. The narrative-trained version successfully learned to align geometric structures with text descriptions; the paper-text version did not.

"We wanted to create an AI model that works broadly across materials," Li said. "That's why we decided to synthesize narratives rather than rely solely on existing papers. We found when we used that narrative dataset, the models are better at predicting the function."

In plain terms: writing fresh, structured descriptions of crystal structures taught the model geometry that reading the existing literature never could.

What changes for materials scientists?

Over the last decade, AI has already reshaped how researchers propose new crystalline structures, typically by treating them as 3D graphs. The new system goes further: give it a natural-language query describing a property, and it identifies crystalline materials with that property — and it outperformed a baseline model in the researchers' tests.

Olivetti places the work in the lineage of field-shifting tools. "There are sometimes these step changes within materials science around algorithmic integration toward useful tools, which the community uses to further advance the field," she said. "This approach helps us see more deeply into three-dimensional structures and link structures to property and processing. I hope this can be one of those step changes."

Li compares the model to AlphaFold, the AI system that links protein structures to function and has accelerated biological research and drug development. The analogy is the mechanism: map structure directly onto function, and downstream work speeds up.

The approach should generalize across many crystalline materials, and the team hopes other materials scientists will adopt it. "Being able to see a three-dimensional crystal and how its atoms are connected is super foundational," Li said. "I think this will be useful to anyone who practices materials science."

The work received partial support from the U.S. Defense Threat Reduction Agency through the University Research Alliance, the U.S. Defense Advanced Research Projects Agency, and the National Research Foundation of Korea.

For now, the findings come from a single paper, and the battery-electrolyte test is one demonstrated use case rather than proof of universal capability. But the direction is clear: give AI a geometric view of matter, and it starts recommending materials that human experts would endorse.