The question

Most useful technology, from batteries to superconductors to solar cells, depends on finding the right material. The trouble is that the space of possible crystals is astronomically large, and testing each one in a lab is slow and expensive. For decades, researchers have used quantum-mechanical simulations to predict which arrangements of atoms are stable, but even those calculations are costly, so only a small slice of the possibilities had ever been explored. The question the DeepMind team set out to answer was simple to state and hard to do: could a machine-learning model reliably predict material stability well enough to expand that catalogue by an order of magnitude?^[1]

What they did

The researchers built GNoME, short for Graph Networks for Materials Exploration, a graph neural network that represents a crystal as atoms connected by bonds and learns to predict its energy.^[1] Lower energy generally means more stable. They generated candidate structures in two ways: a structural pipeline that swaps elements into templates resembling known crystals, and a compositional pipeline that starts from raw chemical formulas with less prior structure. The model's predictions were checked against density functional theory, the standard quantum-mechanical simulation, and the verified results were fed back to retrain the model in repeated rounds. This active-learning loop let GNoME improve as it explored, rather than relying on a fixed training set.^[1]

What they found

GNoME predicted the stability of about 2.2 million structures. Of those, roughly 380,000 (381,000 in the paper) sit on the updated stability frontier as new stable materials, an increase of nearly an order of magnitude over the number previously known.^[1] The final models predicted formation energies to within about 11 meV per atom, and the precision of stable predictions rose to above 80 percent for structural candidates and around 33 percent per 100 trials for composition-only candidates, compared with roughly 1 percent in earlier approaches.^[1] Importantly, this was not purely theoretical: in concurrent work, external labs around the world independently synthesized 736 of the predicted structures, a real-world signal that at least some predictions correspond to makeable materials.^[1]

Why it matters

Screening at this scale changes the economics of materials search. Instead of running an expensive simulation on every candidate, a lab can use a fast model to triage millions of options and spend its limited quantum calculations and bench time on the most promising ones. The stable structures were contributed to public databases, giving the wider community a large pool of candidates to mine for batteries, catalysts and other applications. The paper is also a clear demonstration that scaling up data and model training, the same recipe that transformed language and vision, can extend into the physical sciences.^[1]

What this does not prove

A prediction of stability is not a material in hand. The overwhelming majority of the 380,000 candidates have never been synthesized, and stability on a computed energy landscape does not guarantee that a compound can actually be made, is safe, or does anything useful.^[1] The claim of novelty has also been contested. In a 2024 Perspective, chemists Anthony Cheetham and Ram Seshadri examined a sample of the released structures and reported finding no strikingly novel compounds, arguing that many entries were minor variations of known materials and that weak deduplication may have let structures already present in existing databases be counted as new.^[2] Their broader point is that a useful discovery should be credible, novel and useful all at once, and a raw stability prediction clears only part of that bar. This is a computational result, not experimental confirmation that hundreds of thousands of genuinely new materials exist.

What happens next

The real test is synthesis. Robotic labs and human experimentalists are working through the candidate list to see which predicted crystals can actually be grown and whether their measured properties match the calculations. Sharpening novelty and deduplication checks, so that the count of truly new materials is trustworthy, is an open task the critics have pushed to the front. And the larger question, whether a stability catalogue this size translates into materials people actually use, will take years of lab work to answer.^[1][2]

---

References

1. Merchant A, Batzner S, Schoenholz SS, Aykol M, Cheon G, Cubuk ED. Scaling deep learning for materials discovery. Nature. 2023. doi:10.1038/s41586-023-06735-9
2. Cheetham AK, Seshadri R. Artificial Intelligence Driving Materials Discovery? Perspective on the Article: Scaling Deep Learning for Materials Discovery. Chemistry of Materials. 2024. doi:10.1021/acs.chemmater.4c00643