Your phone, your computer, the solar panels on rooftops, an electric car's battery: all depend on materials with very specific properties. And most of these materials are crystals, meaning atoms arranged in a perfect order that repeats in three dimensions.
Finding new crystals therefore means finding the technologies of tomorrow. A team at MIT has just given AI a new way to search for them.
The problem: AI reads, but does not see
Large AI models, like the ones that run assistants, learned by reading vast quantities of text. That is very useful for answering questions. But matter is not made of words. It is made of atoms arranged in space.
It is a bit like asking someone to choose the best flat by reading only listings, without ever seeing a floor plan. The listings say a lot, but they often leave out the essentials: the shape of the place.
Professor Ju Li, who leads the team, sums it up this way: structure is at the heart of materials science. Understand it, and you can infer a great deal about properties and performance.
The solution: linking shape and words
The researchers created a model that learns to associate two things: the 3D structure of a crystal, and a plain-language description of that crystal, its composition, its properties, how to make it, its possible uses. The principle is similar to the one that lets some AIs link an image to its caption.
Once trained, you can ask it in plain words for a material with a given property, and it will look for the structures that match.
The researchers tested it on a very concrete challenge: finding solid electrolytes for lithium batteries, a route to safer batteries. The model recovered the best known candidates, where an AI fed only on text missed them. Professor Li compares the approach to AlphaFold, the AI that revolutionised the study of proteins.
An old story tells of a man searching for his keys under a streetlight. A passer-by asks whether he is sure he lost them there. "No," he replies, "but this is where the light is." The MIT researchers identified this trap: scientific papers mostly discuss materials that have already been heavily studied, and leave much of the possible in the dark. To avoid it, they trained their model on descriptions generated from databases, which also cover materials no one has yet discussed. That version gave better results than the one fed only on papers.
Why it matters
More than 90% of solid inorganic materials are crystalline. An AI capable of understanding their structure, and of conversing with researchers in plain language, could accelerate the discovery of materials for batteries, solar energy, electronics or green chemistry.
The study was published this summer in the journal Nature Communications, and presented this week by MIT. It remains a step: finding a promising candidate on a computer does not replace laboratory experiments, which alone will tell whether it lives up to its promise.
What to take away
There is much talk of the AI that writes and draws. This one learns to see what cannot be seen with the naked eye: the hidden order of atoms.
It is a useful reminder, after the test showing that AI cannot yet invent on its own: in the hands of researchers, it can already help them search where no one had thought to look.