A flood of new crystals
In late 2023 a pattern-finding program called GNoME, built by Google DeepMind, published a list of about 2.2 million crystal structures that were not in the world's catalogues, and flagged about 380,000 of them as likely to hold together. The team behind it describes the haul as close to 800 years of discovery at the earlier pace. Here is what that means.
Start with something you own
Flip your phone over. The chip inside is a single crystal of silicon, grown to order. The battery stores its charge in a crystal that lets lithium slide in and out like a key in a lock. The magnet holding the shopping list to your fridge is a crystal with its atoms lined up in one direction. The panels on a solar roof catch light with crystals tuned to sunlight.
Nearly every finished thing you can point at leans on a crystal doing a quiet job. So a sudden pile of new crystals is not trivia for a lab shelf; it is a pile of possible future products, waiting to be sorted.
What a crystal is
A crystal is a repeating pattern of atoms, like brickwork. Lay one brick pattern down, repeat it in every direction, and you have a wall; repeat one small cluster of atoms in every direction and you have a crystal. Salt, diamond, the quartz in a watch: each is one small arrangement, copied over and over, trillions of times.
The pattern is the whole game. Carbon stacked one way is soft grey pencil lead; the same carbon in a tighter pattern is diamond. Change the pattern and you change the material. Glass, by contrast, has no long-range repeat. Each atom still holds its immediate neighbours at set distances, but the pattern loses step within a nanometre or so, a millionth of a millimetre, and never comes back: like bricks tipped out of a truck landing on their faces, without a wall ever forming.
The old way: slow cooking and luck
For most of the last century, finding a new material meant mixing powders, baking them in a furnace for days, grinding the result, baking again, and then measuring whether anything new had formed. Months of careful work could end in a grey lump of nothing much. Good materials also arrived by accident, noticed by someone paying attention to a failed batch.
The recipe book grew, but slowly. Averaged over the long run, the world's list of crystals known to be stable, meaning they hold their pattern rather than falling apart into simpler stuff, grew by a few hundred entries a year.
What GNoME is
GNoME reads a crystal the way you might read a dot-to-dot puzzle: each atom is a dot, each bond between atoms is a line. Under the hood it is a graph neural network, which is a program that learns from worked examples which arrangements of dots and lines tend to hold together and which tend to fall apart.
Fed the tens of thousands of crystals already in the catalogues, it began proposing new ones: sometimes a known pattern with one kind of atom swapped for another, sometimes a wilder arrangement no bench had tried. Most proposals fail the check, and that is fine, because it can make millions of them.
The loop
Propose
GNoME proposes a batch of new atomic arrangements, from careful swaps to long shots.
Check
Physics maths grades each proposal: would this arrangement hold its shape, or collapse into something simpler?
Learn
The graded proposals, passes and fails alike, go back into the program's training pile.
Propose better
Each round comes back sharper, and GNoME ran six of them. Of the proposals made by swapping atoms into a known pattern, fewer than 6 in 100 survived the check in the first round and more than 80 in 100 in the last. Proposals built from a formula alone, with no pattern to copy, went from under 3 in 100 to above 33 in 100.
The numbers, held one at a time
Numbers this big slide off the mind, so give each one a picture. About 2.2 million predicted crystals is close to one for every person in Perth, which passed 2.4 million in 2025. The stable short-list, about 380,000, would fill the MCG (the Melbourne Cricket Ground, about 100,000 seats) close to four times over.
Or try it as a job. If you looked at one predicted crystal a minute, eight hours a day, seven days a week, the full 2.2 million would keep you busy for about 12 and a half years. The stable short-list alone would take you a bit over two years.
The team describes the release as close to 800 years of discovery at the earlier pace. Count back 800 years from 2023 and you land in the 1220s, when the Magna Carta was newly signed. Before the release, the catalogues held about 48,000 crystals known to be stable; with it, about 421,000, which is nearly nine times the earlier count.
Fresh patterns proposed by GNoME and graded by the physics maths.
Arrangements the maths says would hold their shape: the short-list worth trying at a bench.
Made by lab teams around the world, each working on their own, while GNoME was still running.
Made by the A-Lab in 17 days, from targets drawn from the same shared pool of predictions.
Bar lengths are compressed so the small counts stay visible; read the numbers for the real drop.
Inside the pile are families worth naming. About 52,000 of the predictions are layered materials similar to graphene, the carbon sheet one atom thick; that is close to a full house at Suncorp Stadium in Brisbane, one candidate per seat. Another 528 look like lithium-ion conductors, solids that could let lithium, the moving part of a battery, slip through; the team reports that is about 25 times as many candidates as an earlier search of the same kind turned up.
How fast is that, really?
Take the team's own comparison at face value: about 380,000 stable newcomers in one release, against roughly 800 years at the earlier pace. That works out to an earlier pace of a few hundred a year. Drag the slider to see how long the old road would have been.
The robot kitchen
Predicting a crystal is one thing; making it is another. In a room at Lawrence Berkeley National Laboratory in California sits the A-Lab: robotic arms that weigh out powders, load furnaces, bake, cool, and then read their own results. When a bake misses, the system rewrites its own recipe and tries again, drawing its cooking ideas from thousands of published methods.
Over 17 days of running on its own, the A-Lab made 41 of the 58 target crystals it attempted, working from the same shared pool of predictions. That is better than two new materials a day. A human team can spend months on a single one, and still miss.
What it could mean
Everything below is about what could follow, not what has been built. A predicted crystal is a candidate, not a product.
Batteries
The 528 candidate lithium conductors could lead to solid-state batteries: batteries with no flammable liquid inside, which could hold more charge and shrug off knocks. Any one of them would still need years of testing before it reached a car or a phone.
Solar
New light-catching crystals could lift how much of the day's sun a rooftop panel turns into power, and the layered candidates could suit panels that are lighter and cheaper to make. Which ones survive rain, heat and twenty summers is the open question.
Computing
The team points to candidate superconductors, materials that could carry current with no energy lost as heat. If even a few worked at friendlier temperatures, chips and power lines could run cooler and waste less along the way.
Browse it yourself
The predictions were handed to the Materials Project, a free public database run out of Berkeley that lab teams and students anywhere can search. Type in a pair of elements and see what the maths says they could build together. No fee, no special kit; a browser is enough.
Limits
A prediction is a computed judgement, not a sample in a jar. The stability check is done with physics maths at idealised conditions, colder and cleaner than any real bench; how a crystal behaves warm, damp and squeezed is a separate question, answered only by making it.
Stable does not mean useful, cheap or easy to make. Many of the 380,000 may resist every furnace recipe thrown at them, and how many of the rest are worth a factory is a question only a furnace answers. So far 777 have been through one: 736 made by lab teams around the world plus the A-Lab's 41, a first taste and not a verdict on the rest.
The headline growth is from about 48,000 known stable crystals to about 421,000. That is nearly nine times the earlier count, which the team rounds off to about ten times over. The 800 years figure is the team's own comparison against the earlier pace, and it depends on how that pace is measured.
Sources
- Merchant et al., "Scaling deep learning for materials discovery", Nature, November 2023 · nature.com/articles/s41586-023-06735-9
- Szymanski et al., "An autonomous laboratory for the accelerated synthesis of inorganic materials", Nature, November 2023 · nature.com/articles/s41586-023-06734-w
- "Millions of new materials discovered with deep learning", Google DeepMind, November 2023 · deepmind.google/discover/blog/millions-of-new-materials-discovered-with-deep-learning
- "Google DeepMind adds nearly 400,000 new compounds to Berkeley Lab's Materials Project", Lawrence Berkeley National Laboratory, November 2023 · newscenter.lbl.gov/2023/11/29/google-deepmind-new-compounds-materials-project
- The Materials Project, free public database · materialsproject.org