Learned models can explore candidate structures beyond conventional computation and human-chosen searches.
Executive brief
The workflow gain is candidate-space expansion; governance requires keeping predicted stability, successful synthesis, and useful material performance as separate gates.
AI value · New predicted stable crystal structures
Verified2.2 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.
Before
Select candidate structures and run high-cost stability calculations. → Choose a small subset for synthesis.
After
Predict energies and identify candidate stable structures at large scale. → Validate candidates with DFT, database comparison, and synthesis.
Human boundary
GNoME ranks predicted stability; scientists and higher-fidelity computation determine whether candidates are credible and worth synthesis.
Why it matters
Learned models can explore candidate structures beyond conventional computation and human-chosen searches.
Before
How the work ran before the change.
Computational materials scientist
Select candidate structures and run high-cost stability calculations.
ControlHuman search strategy
Experimental scientist
Choose a small subset for synthesis.
ControlExpert judgment
What changed
Learned models can explore candidate structures beyond conventional computation and human-chosen searches.
Decision rightSelection moves from a fixed rule to the model
After
How the same work runs now.
GNoME graph networks
Predict energies and identify candidate stable structures at large scale.
ControlModel hit-rate and convex-hull criteria
Computational and experimental scientists
Validate candidates with DFT, database comparison, and synthesis.
ControlHumans retain validation and usefulness decisions
Exception path
Predictions can be displaced by future discoveries, fail higher-fidelity calculations, or prove unsynthesizable; those cases remain outside validated material knowledge.
Decision authority
GNoME ranks predicted stability; scientists and higher-fidelity computation determine whether candidates are credible and worth synthesis.
| # | Actor | Action | Control |
|---|---|---|---|
| 01 | Computational materials scientist | Select candidate structures and run high-cost stability calculations. | Human search strategy |
| 02 | Experimental scientist | Choose a small subset for synthesis. | Expert judgment |
| # | Actor | Action | Control |
|---|---|---|---|
| 01 | GNoME graph networks | Predict energies and identify candidate stable structures at large scale. | Model hit-rate and convex-hull criteria |
| 02 | Computational and experimental scientists | Validate candidates with DFT, database comparison, and synthesis. | Humans retain validation and usefulness decisions |
Work that left the path
Human role before
Scientists selected and screened a much smaller candidate space.
Human role after
Scientists validate, prioritize, synthesize, and assess use from a vastly expanded model-generated candidate set.
AI role
Graph neural networks predicting crystal energy and stability to expand the candidate search space.
About 48,000 stable crystals in external datasets by 20232.2 million structures stable relative to prior work; 381,000 entries on the updated convex hull
Research through July 2023, published 2023-11-29 · Large-scale computational exploration with about 216,000 consistent DFT calculations used for comparison
Predicted thermodynamic stability is not experimental synthesis or practical utility; future discoveries can displace convex-hull entries.
Anti-pattern
Calling 381,000 predictions newly manufactured materials.
Questions
Portability conditions
Reputation risk
medium