What happened?
In November 2023, Google DeepMind described GNoME (Graph Networks for Materials Exploration), a deep learning tool that predicts whether a crystal structure is stable DeepMind blog ↗. The team reported 2.2 million new crystal structures, of which about 381,000 were judged stable by computer calculations Nature paper ↗. Stable here means the arrangement of atoms should not break down into other, simpler compounds.
Before this work, the team counted about 48,000 known stable crystals in the databases they started from Nature paper ↗. The paper reported that 736 of GNoME's stable structures had already been made by other scientists in separate work, found by matching them against a large database of measured crystals Nature paper ↗. A companion paper from Berkeley Lab described a robotic lab, called A-Lab, that made more than 41 new materials guided by AI predictions DeepMind blog ↗. DeepMind released the 380,000 most stable predictions to the Materials Project, a free public database, and posted the data on GitHub DeepMind blog ↗.
In April 2024, chemists Anthony Cheetham and Ram Seshadri of UC Santa Barbara published a critique in the journal Chemistry of Materials Chem. Mater. critique ↗. They checked 10 randomly chosen compounds from the stable list and found that all 10 resembled structures already in the main crystal database Chem. Mater. critique ↗. They wrote that they had "yet to find any strikingly novel compounds" and found "scant evidence" of compounds that are new, believable and useful at the same time Chem. Mater. critique ↗. DeepMind told The Register it stands by all claims in the paper The Register ↗.
What do the critics' three tests mean?
Novelty
A compound should not be a small tweak of a known one. The critics said many predictions are "trivial adaptations of known materials" Chem. Mater. critique ↗.
Credibility
A compound should be something a lab could plausibly make. The critics flagged atom orderings unlikely at real synthesis temperatures, and about 18,000 entries using radioactive elements Chem. Mater. critique ↗.
Utility
A compound should do something useful. The critics noted that none of the listed compounds had shown a working function, so they called them proposed compounds rather than materials Chem. Mater. critique ↗.
A much larger map of possibly stable crystals now sits in a free database, which can help experimenters decide where to look. The debate shows that a big list is a starting point, not a finished discovery.
Leapscope interpretation of the reported result.How did AI help?
GNoME is a graph neural network, a type of AI that treats a crystal as a network of atoms and the links between them DeepMind blog ↗. It proposed candidate structures and predicted their stability, and standard physics calculations then checked the most promising ones; the results fed back to improve the model in rounds Nature paper ↗. The paper reports that structure based predictions were right more than 80% of the time, compared with about 1% in earlier work Nature paper ↗. Humans designed the system, chose the checks and wrote the paper.
Figures from the Nature paper Nature paper ↗. The 736 is a match to existing lab records, not new synthesis by DeepMind.
The main limit is that stability on a computer is not the same as a useful material in hand. Cheetham and Seshadri argued that the methods "appear to hold promise" but need more input from experts in making and measuring crystals The Register ↗. They also said the patterns of crystal symmetry in GNoME's list differ sharply from those seen in real crystals, which suggests some predicted structures are artifacts of the method Chem. Mater. critique ↗.
Which fields could this affect?
The immediate value is a large public list of candidates; uses in products are possible later. These connections are our assessment.
Materials science research
Researchers can search the released predictions in the Materials Project DeepMind blog ↗. Critics advise filtering them against existing literature first Chem. Mater. critique ↗.
Explore scienceBatteries
DeepMind highlighted 528 possible lithium ion conductors among the predictions DeepMind blog ↗. None has been shown to work in a battery in these papers.
Explore scienceAutomated labs
A-Lab showed robots can try to make AI suggested compounds DeepMind blog ↗. How often such attempts yield genuinely new materials is still debated.
Explore scienceElectronics and chips
DeepMind pointed to about 52,000 layered compounds similar to graphene DeepMind blog ↗. Turning any of them into a device would need synthesis, testing and engineering not shown here.
Explore softwareWhat has been checked?
The core result is a peer reviewed Nature paper, followed by a peer reviewed critique in Chemistry of Materials. Leapscope reviewed these sources; we did not repeat the experiments or calculations.
Shown so far
- GNoME produced 2.2 million candidate structures, 381,000 of them predicted stable by standard calculations Nature paper ↗.
- 736 of the stable predictions match crystals that other labs had already made Nature paper ↗.
- An outside check of 10 random entries found all 10 resembled crystals already on record Chem. Mater. critique ↗.
Still unknown
- What share of the 381,000 are genuinely new rather than small variations of known compounds.
- How many can actually be made in a lab under normal conditions.
- Whether any will prove useful in a real device or product.
Evidence status: Published research. Stage: Claim. Mostly computer predictions. Only a small share of the materials has been made in a lab.
From a list to a material
This is our suggested way to follow the story, not a promised timetable.
Can I use it today?
Yes, for researchers: the predicted structures are free in the Materials Project and on GitHub DeepMind blog ↗. For everyone else, there is no product. These are computer predictions, not materials you can buy.
A few things you might be wondering
Did AI really discover millions of new materials?
It predicted millions of crystal structures, and about 381,000 were judged stable by calculation Nature paper ↗. Outside chemists argue most are compounds, not materials with a shown use, and that many are close to known ones Chem. Mater. critique ↗.
Has anyone made these materials?
736 match crystals that other labs had already made Nature paper ↗. A robotic lab at Berkeley made more than 41 new materials guided by AI predictions DeepMind blog ↗.
Did Google respond to the criticism?
Yes. A spokesperson said the company stands by all claims in the paper and that hundreds of predictions had been independently made by scientists The Register ↗.
Go straight to the sources
Checked Oct 8, 2026. The first source is the original announcement or research. Later sources add independent context; background pages do not validate the result on their own.
01DeepMind's summary of GNoME, its numbers and the public data release.
The research paper with methods, accuracy figures and the 736 experimental matches.
Cheetham and Seshadri's critique of how novel, credible and useful the predicted compounds are.
Reports the critique, interviews Cheetham and includes DeepMind's response.