In November last year, artificial intelligence (AI) Graph Networks for Materials Exploration (GNoME) claimed to have successfully discovered 2.2 million new crystal structures with 384,870 of them stable. Deepmind said GNoME made discoveries equivalent to 800 years of manual study.
But last week two researchers from the University of California, Santa Barbara published a research paper in the journal Chemical Materials refuting Deepmind's claims. They tested a random sample of 380,00 stable crystal structures found by GNoME and found that none passed the definitions of "useful", "new" and "reliable". Because it does not pass these three standards, all of these crystal structures cannot be considered new materials as declared by Deepmind.
However, Deepmind still defends the GNoME findings because the researchers who refuted them only took a small sample and not all 384,870 stable structures.