Google DeepMind's GNoME finds 2.2 million new crystals

Background

Finding a new material is largely a search problem. A crystal is a solid whose atoms sit in a repeating arrangement, and even for a handful of elements the number of plausible arrangements is enormous. Most of them do not survive: the atoms rearrange, or the structure comes apart. For that reason, computational screening has become a standard first step — using the laws of physics and chemistry to estimate which candidate structures are likely to be stable before anyone tries to make them in a laboratory.

On 29 November 2023, Google DeepMind announced GNoME, a deep-learning system for discovering new crystal structures, together with a paper in Nature ^[1]. The announcement came with two Nature papers: one from Google DeepMind and one led by researchers at Berkeley Lab ^[1].

What was announced

The headline figure was 2.2 million new crystals, of which 380,000 were predicted to be among the most stable — the candidates most likely to be worth attempting ^[1]. Stability here is a prediction rather than a guarantee: it means the model expects a structure to hold together instead of decomposing into something else.

The scale is easiest to see in a comparison. Counting only materials identified through computational methods, the number of known stable materials had stood at about 48,000; with GNoME it rose to 421,000 ^[1].

Key details

A prediction is only useful if it survives contact with an experiment, and two kinds of check were reported. First, 736 of the new materials had been independently created by external researchers, so those particular predictions were confirmed by people outside the project ^[1]. Second, an autonomous lab at Berkeley Lab — a laboratory where robotic equipment carries out experiments with little direct human handling — synthesised more than 41 new materials using the predictions ^[1].

The stable structures were to be added to the Materials Project, a shared database of materials information that researchers use to look up and compare structures ^[1]. Putting the candidates into a common database makes them available to other groups, who can then choose which ones to try to make.

Why it matters

The bottleneck in materials discovery has two halves. One is coming up with good candidates; the other is testing them. A list of millions of structures is only as valuable as the number of them that can actually be made and measured, and synthesis is slow, expensive and often fails. The reported numbers speak to that gap: hundreds of the predictions were reproduced by outside researchers, and dozens more were made in an automated laboratory ^[1].

That is why the pairing of a large generative search with automated synthesis is the part worth watching. The model widens the set of candidates; robotic experiments shrink the time needed to check them. Neither half is new on its own, but the two together describe a loop in which predictions are made, tested and fed back.

There is a further, quieter consequence. The count of known stable materials rising from about 48,000 to 421,000 changes what a materials database contains ^[1]. Reference collections that researchers use to look up whether a compound is known, and what its properties are, become a different kind of resource when most of their entries have been predicted rather than measured. The Materials Project entry for a structure becomes a pointer to something not yet made.

What the announcement does not settle is how many of the 2.2 million crystals will ever exist as real samples. The 380,000 most-stable candidates are predictions, and the 736 independently created materials and the 41 synthesised in the autonomous lab are the evidence offered so far ^[1]. Turning the rest into knowledge is laboratory work, one synthesis at a time.