# Study proposes stricter tests for novelty claims in AI materials discovery

> Researchers at the University of Liverpool and University College London used continuous mathematical invariants to compare crystal structures. Their preprint examines two GNoME dataset releases and argues that duplication checks are essential before claiming new materials.

- Canonical URL: https://newsverum.com/en/news/materials-ai-benchmarking-gnome-duplicates-2026
- Publisher: Newsverum
- Section: Technology
- Country: United Kingdom
- Published: 2026-10-13T00:51:40+00:00
- Updated: 2026-10-13T00:51:40+00:00
- Event / data period: 2026-10-09
- Other editions: [Türkçe](https://newsverum.com/tr/news/materials-ai-benchmarking-gnome-duplicates-2026.md), [العربية](https://newsverum.com/ar/news/materials-ai-benchmarking-gnome-duplicates-2026.md), [Français](https://newsverum.com/fr/news/materials-ai-benchmarking-gnome-duplicates-2026.md), [Español](https://newsverum.com/es/news/materials-ai-benchmarking-gnome-duplicates-2026.md), [Deutsch](https://newsverum.com/de/news/materials-ai-benchmarking-gnome-duplicates-2026.md), [Português](https://newsverum.com/pt/news/materials-ai-benchmarking-gnome-duplicates-2026.md)

The team compared crystals in stages, using distributions of interatomic distances and stronger structural descriptors derived from them. The paper explains why a different chemical formula or CIF file alone does not establish a structurally new material.

The authors report 554,054 entries in GNoME2, of which 301,413 also appear in GNoME1, and compare the structures across the two releases with invariant-based methods. This is a database audit, not a count of newly made experimental materials. They propose separate checks for novelty, stability and synthesizability before automated discovery systems make claims.

## Data

### Reported study result

Database-entry total reported by the authors; it is not a count of newly made experimental materials.

Methodology: The paper reports 554,054 GNoME2 entries and separately compares structural duplicates with GNoME1.

| Reported quantity | Value | Study-specific scope |
|---|---|---|
| Entries in the GNoME2 dataset examined by the study | 554054 | Database-entry total reported by the authors; it is not a count of newly made experimental materials. |

Sources: <https://arxiv.org/abs/2610.13885> · Download data: [JSON](https://newsverum.com/api/v1/articles/materials-ai-benchmarking-gnome-duplicates-2026/data/study-result?lang=en&format=json) · [CSV](https://newsverum.com/api/v1/articles/materials-ai-benchmarking-gnome-duplicates-2026/data/study-result?lang=en&format=csv)

## Sources

- [arXiv](<https://arxiv.org/abs/2610.13885>) (Source checked: 2026-10-13; Published: 2026-10-09)

---

Cite the canonical URL above. [Editorial standards and corrections](https://newsverum.com/en/editorial)