GNoME predicts 2.2 million crystal structures, and the claim is disputed

Google DeepMind published GNoME in Nature on 29 November 2023, reporting 2.2 million predicted crystal structures of which 381,000 were judged stable, and describing this as an order of magnitude expansion in the stable materials known. A companion paper from Berkeley Lab reported a robotic laboratory synthesising 41 of the predicted compounds.

Background

A new inorganic crystal takes months of laboratory work to make and characterise, and most attempts fail because the compound is not stable enough to exist. Screening candidates computationally had been tried for years, with a success rate around one per cent.

The Materials Project at Berkeley Lab had assembled a database of roughly 200,000 known materials over a decade, which is the scale the field was working at.

What happened

Google DeepMind published GNoME in Nature on 29 November 2023. It predicts 2.2 million crystal structures and judges 381,000 of them stable, which DeepMind described as multiplying the technologically viable materials known to humanity and as equivalent to nearly 800 years of knowledge.

The 380,000 most stable were contributed to the Materials Project. A companion paper from Berkeley Lab described A-Lab, a robotic facility that used the predictions to synthesise 41 new compounds with little human involvement.

What followed

In April 2024 Anthony Cheetham and Ram Seshadri published an analysis of the claim and found, in their words, scant evidence of compounds meeting all three of novelty, credibility and utility. Many of the predicted structures substitute one element for another in arrangements already known, which is a reasonable thing for a model to produce and not what most chemists mean by a new material.

They did not dismiss the method. Their argument was that the work needed expertise in synthesis and crystallography that it had not drawn on, and that counting predictions is not the same as counting discoveries.

The companion synthesis paper drew its own criticism about whether the compounds produced had been correctly identified.

Both things are true at once. A model generated candidate structures at a scale no previous method approached, and whether that constitutes discovery depends on a question the paper did not settle: how many can actually be made.

Why it mattered

Five months later Cheetham and Seshadri published an analysis finding scant evidence of compounds that were at once novel, credible and useful, noting that many predictions substitute elements into structures already known. The entry records the result and the dispute together, because a prediction is not a discovery until somebody makes the thing.

Sources

• Scaling deep learning for materials discovery. nature.com. Primary source
• Millions of new materials discovered with deep learning. deepmind.google. Official
• Google DeepMind Adds Nearly 400,000 New Compounds to Berkeley Lab's Materials Project. lbl.gov. Secondary

Cite this page

AI Achievements. (2023). GNoME predicts 2.2 million crystal structures, and the claim is disputed. Retrieved 2026-08-29, from https://achievements.ai/milestone/gnome-materials-discovery

@misc{achievements_gnome_materials_discovery,
  title  = {GNoME predicts 2.2 million crystal structures, and the claim is disputed},
  author = {{AI Achievements}},
  year   = {2023},
  url    = {https://achievements.ai/milestone/gnome-materials-discovery}
}