GNoME
Graph Networks for Materials Exploration
Graph neural networks and symmetry-aware structural generators propose inorganic crystals, and Density Functional Theory calculations label their stability to retrain the models across active-learning rounds.
The loop
- Proposes
Candidate inorganic crystal structures generated via elemental substitution, symmetry-aware partial random structures and GNN energy filtering
- Judged by
Ab initio Density Functional Theory (DFT) relaxation in VASP computing formation energy relative to the competing-phase convex hull
- Steers
DFT energies and relaxed geometries expand the training set and retrain the GNoME graph neural networks for the next discovery round
Headline result
In Nature (2023), GNoME identified 2.2 million crystal structures below the previous convex hull of known materials, including 381,000 structures on the updated convex hull (nearly an order of magnitude more than the roughly 48,000 stable inorganic crystals previously catalogued). Active learning raised the DFT stability hit rate from under 10% to over 80% for compositional pipeline candidates and from 6% to 33% for structural candidates.
What this does not prove
DFT stability is calculated at 0 K for ordered unit cells and does not establish room-temperature phase stability, configurational entropy or practical lab synthesis. A 2024 Chemistry of Materials assessment by Cheetham and Seshadri argued that many predicted compounds are routine ionic or intermetallic substitutions with limited novelty or proven utility.
How it generalizes the harness
- A domain authority judges
- Feedback retrains the generator
- Search over a population
- The generator is not an LLM
- The output outlives the model
- Multi-fidelity screening ladder
The generator is a graph neural network pipeline rather than an LLM, VASP quantum-mechanical DFT calculations judge thermodynamic stability, and the discovered crystal catalogue is usable without the model.
Classification
| What it proposes | Crystal structure |
|---|---|
| Generative model | Message-passing Graph Neural Networks (GNoME) paired with structural and compositional candidate pipelines (not an LLM) |
| Algorithmic authority | Density Functional Theory (DFT, VASP) structural relaxation and convex-hull phase stability evaluation |
| Feedback | New labeled data |
| What gets updated | Model weights and training data |
| When it runs | During training |
| Validity | Approximate physics (simulation, not lab results) |
| Evidence | Experiments reported |
| Lineage | Original system |