---
title: 'Google DeepMind: GNoME graph-network crystal stability discovery'
slug: deepmind-gnome-materials-discovery
stable_id: 88c7ce548e467e12
company: Google DeepMind
function_code: healthcare_screening
pattern_codes:
  - continuous_decisioning
evidence_strength: verified
publication_tier: showcase
freshness: current
reviewed_at: '2026-08-23'
updated_at: '2026-08-22'
source_quality_summary: 1 peer reviewed; publication outcomes are verified.
caveat_summary: >-
  Predicted thermodynamic stability is not experimental synthesis or practical utility;
  future discoveries can displace convex-hull entries.
collections: []
bundle_version: 1.0.0
bundle_fingerprint: sha256:c23c6cc2b88153d008ea8fda928f632ce0011fc2d4c5036672a16e5d895bab93
canonical_url: https://brianletort.ai/transformations/deepmind-gnome-materials-discovery
---

# Google DeepMind: GNoME graph-network crystal stability discovery

Learned models can explore candidate structures beyond conventional computation and human-chosen searches.

Function: Healthcare screening. Patterns: Continuous decisioning. Evidence: verified.

Freshness: current. Reviewed: 2026-08-23. Updated: 2026-08-22.


Source quality: 1 peer reviewed; publication outcomes are verified.

## Before

1. **Computational materials scientist** — Select candidate structures and run high-cost stability calculations. (control: Human search strategy)
2. **Experimental scientist** — Choose a small subset for synthesis. (control: Expert judgment)

## After

1. **GNoME graph networks** — Predict energies and identify candidate stable structures at large scale. (control: Model hit-rate and convex-hull criteria)
2. **Computational and experimental scientists** — Validate candidates with DFT, database comparison, and synthesis. (control: Humans retain validation and usefulness decisions)

## Decision rights

GNoME ranks predicted stability; scientists and higher-fidelity computation determine whether candidates are credible and worth synthesis.

## Exception path

Predictions can be displaced by future discoveries, fail higher-fidelity calculations, or prove unsynthesizable; those cases remain outside validated material knowledge.

## Outcomes

- **New predicted stable crystal structures** (verified): About 48,000 stable crystals in external datasets by 2023 → 2.2 million structures stable relative to prior work; 381,000 entries on the updated convex hull. Predicted thermodynamic stability is not experimental synthesis or practical utility; future discoveries can displace convex-hull entries.

## Executive lesson

The workflow gain is candidate-space expansion; governance requires keeping predicted stability, successful synthesis, and useful material performance as separate gates.

## Anti-pattern

Calling 381,000 predictions newly manufactured materials.

## Questions for leaders

- What fraction crosses each validation gate?
- Is the bottleneck discovery, synthesis, or application testing?

## Sources

- [Scaling deep learning for materials discovery](https://www.nature.com/articles/s41586-023-06735-9) — Nature