MIT model learns to read crystal structures and plain text together

Researchers at MIT have built a multimodal AI model for crystalline materials that pairs a material's 3D atomic structure with a natural-language description of it. The work was reported on 19 October by Phys.org, republished from MIT News, and is published in *Nature Communications* as "Scaling multimodal materials representation learning with synthetic narratives".

What they built

Most AI models used in chemistry work from text alone. This one learns from two inputs at once: the geometry of a crystal and a written account of it, covering composition, properties, processing and applications. The team generated those descriptions synthetically and trained the model with contrastive learning, an approach that teaches a model to match the right structure with the right text.

The researchers report that the model did better than text-only models at identifying materials, and that it learned geometric representations text-only models could not. The article does not publish performance figures.

Crystalline materials account for more than 90% of the solid materials on Earth, according to the article, which gives a sense of how much ground a better representation could cover.

Who's behind it

The first author is Yang Jeong Park, a former MIT postdoc. The team includes Ju Li, the Carl Richard Soderberg Professor of Power Engineering in MIT's Department of Materials Science and Engineering, and Elsa Olivetti, the Jerry McAfee (1940) Professor in Engineering in the same department, along with Mayank Kumaran, a recent computer engineering graduate of the University of Illinois Urbana-Champaign, and Chia-Wei Hsu.

Why it matters for materials informatics

Commercial platforms in this market increasingly promise that a chemist can describe what they need in plain language and get candidate materials back. A model that ties structure and text together is a step towards that kind of interface. Turning research like this into a tool that works on a company's own data is a separate job, and that is where software vendors come in.

Money is flowing into the category. Earlier in October, Holyvolt raised $45 million for an approach that connects AI models to automated labs that build and test real battery cells, and University of Liverpool spin-out Matter Machine raised £4 million to find and optimise inorganic crystalline materials. Tech Funding News also points to CuspAI's $450 million Series B in July and Periodic Labs' $300 million seed round.

The bigger picture

The common thread is that models alone are not enough. Holyvolt founder and CEO Mathias Ingvarsson put it simply in the company's funding announcement: "only real-world experiments can tell you whether it works". In our view, the vendors that win chemicals and materials customers will be the ones that join models like MIT's to experimental data and the workflow software R&D teams already use.