GNoME and the materials-discovery claim

Assembled from 19 blocks · 2 evidence · 26 related

1. Story
2. Google DeepMind's Graph Networks for Materials Exploration, announced in Nature in late 2023 with the claim that it had found 2.2 million new crystal structures, of which about 380,000 were predicted stable — described publicly as the equivalent of roughly 800 years of accumulated knowledge. It is included in this Journey because of what happened next. In April 2024, writing in Chemistry of Materials, Anthony Cheetham and Ram Seshadri of the University of California, Santa Barbara took a random sample of the proposed structures and reported that none of the sampled entries met a three-part test of being credible, useful and novel, that many were compositional substitutions on already known compounds, and that the outputs should be described as predicted crystalline inorganic compounds rather than as materials. The disagreement is not about whether the model produced predictions; it is about whether a computationally predicted stable structure constitutes a discovery. Neither position has been settled by synthesis at scale, and that is the honest state of the case.
3. Knowledge
4. GNoME and the materials-discovery claim
5. Google DeepMind
6. Connections
7. Google DeepMind
8. GNoME is announced as 2.2 million new crystal structures
9. Crystallographers dispute the discovery claim
10. AlphaFold
11. GNoME and the materials-discovery claim
12. London
13. AlphaGo defeats Lee Sedol in Seoul
14. AlphaFold2 is assessed blind at CASP14
15. GNoME is announced as 2.2 million new crystal structures
16. GenCast is published with a like-for-like comparison against an operational ensemble
17. Evidence
18. Graph Networks for Materials Exploration produced 2.2 million predicted new crystal structures, of which about 380,000 were predicted stable, presented publicly as equivalent to roughly 800 years of accumulated knowledge. V55 verification basis: neither the paper nor the announcement was fetched. The 2.2 million and roughly 380,000 figures, the "800 years" framing and the Nature publication in 2023 were carried in search results attributing them to the paper and to DeepMind's announcement; the Nature volume, article number and page range were NOT retrieved and are deliberately omitted, as is the full author list.
19. Examining a random sample of the roughly 380,000 structures GNoME proposed as stable, the authors report finding none that met a three-part test of being credible, useful and novel; that many entries were compositional modifications of already known compounds; and that the outputs should be described as predicted crystalline inorganic compounds rather than by the more generic label "material". V55 verification basis: the article was not fetched — pubs.acs.org is blocked to this session — but search results carried the title, the DOI 10.1021/acs.chemmater.4c00643, the authors and their institution, and the substance of the three-part test and the conclusion, attributing them to this perspective. The size and selection method of their random sample were NOT retrieved and are not asserted.