MIT teaches AI to see how crystals are built, and that helps the search for new materials

Almost everything that is solid, from salt and metal to the stones under our feet, is made of crystals. According to MIT, that is more than 90 per cent of all solid substances on Earth. In a crystal, the atoms lie in a fixed, repeating pattern, like the boxes on a neatly stacked pallet. That pattern largely determines how strong, conductive or useful a material is.

Anyone looking for a new material, for example for a better battery, therefore wants to know which structure goes with which property. AI can help with that search. But there is a problem: large language models, such as the models behind AI chatbots, learn mainly from text. The structure of crystals, by contrast, is stored in special computer files full of coordinates, which such models hardly understand.

Researchers at the Massachusetts Institute of Technology (MIT) have now found a solution. They described their work in the journal Nature Communications. The team is led by professors Ju Li and Elsa Olivetti of the Department of Materials Science. The first author is Yang Jeong Park, formerly a researcher at MIT.

Crystals with a story

The idea: give the model both the 3D structure and a description in plain language for each crystal. The researchers did not write those descriptions by hand. They had them generated from large databases of calculated and measured crystals. Each description tells a small story about the material: what it is made of, how the atoms are arranged, what properties it has, how to make it and what it could be used for.

Why not simply use texts from scientific papers? Those are mainly about a few popular kinds of crystals. A model that learns only from them gets a distorted picture of the world of materials. The generated descriptions cover a much broader group of crystals.

Linking structure and language

The model learns with a method called contrastive learning. It is shown pairs again and again: a crystal with the right description, and a crystal with a wrong one. In this way it learns which structure goes with which words.

The researchers compared two versions. One learned from the generated descriptions, the other only from texts from scientific papers. Only the first version really learned to connect the shape of a crystal with the words that describe it.

Better candidates for batteries

The team then tested whether the model can also search. You ask a question in plain language about a property, and the model points to crystals that match it. It did this better than a comparable model without this approach, and it found better material designs than models that know only text.

One example is solid electrolytes for lithium-ion batteries. The electrolyte is the part of a battery through which the charged particles travel. A solid instead of a liquid electrolyte can make batteries safer. Among the best candidates the model proposed were materials that other researchers had independently already found promising. That is a good sign that the model makes sensible choices.

We wrote earlier about Japanese researchers who also had an AI model look at the electrons in a crystal. Both teams are looking for ways to give AI more physical insight, so that it learns more from the same data.

What comes next

Li compares the work to AlphaFold, the AI system that linked the shape of proteins to how they work. According to him, you can deduce a lot about a material once you know its structure. Olivetti speaks of a possible big step forward, because the model connects the 3D structure with properties and production.

The approach works for crystals. Materials with many irregularities are not yet covered. The researchers hope that other materials scientists will use their method to work faster. The next step is for such models not only to point out materials, but also to propose new designs on request.

What does this mean for the road to the singularity?

Language models have a lot of knowledge from all fields at once, but until now they saw the physical world mainly through text. This research gives them a new sense: they learn to read the structure of matter itself. That is exactly what is needed to turn AI from a good reader into a real discoverer.

What makes this powerful is the combination. A model that understands both the structure of a crystal and the language of biologists, chemists and engineers can make connections that a specialist in a single field would miss. Think of a material from chip technology that also makes a battery better. Better batteries, chips and solar cells are in turn the building blocks of faster computers, and therefore of better AI. In this way progress reinforces itself.

What is the technological singularity? Read the explanation →