Key Insights
- A new paper details a large amount of very similar entries both within a crystallographic database produced by Google DeepMind’s GNoME and between GNoME’s database and other repositories.
- The paper follows up on and furthers previous evidence suggesting that GNoME’s artificial intelligence–enabled predictions of the structures of materials lack novelty and uniqueness.
- Researchers suggest that determining the novelty of materials that have been hypothesized by AI or produced physically requires careful analysis and vetting to further genuine materials discovery and avoid repeat entries in databases.
Outside researchers are bringing up more questions about a 2023 paper from researchers at Google DeepMind. The outside researchers say the previous study’s database of artificial intelligence–generated crystal structures is littered with duplicates of known and hypothesized crystal structures found in other databases.
That previous paper, published in the high-profile journal Nature, and its authors’ commentary, released at the same time, included big claims. Specifically, the team claimed that their AI tool—dubbed graph networks for materials exploration (GNoME)—had predicted more than 380,000 previously unknown stable compounds. Those compounds, they suggested, may yield promising materials for use in electronic devices and batteries. The team’s commentary claimed that their findings were equivalent to “nearly 800 years’ worth of knowledge.”
