AI Tools for Materials Science
AI models are compressing the materials discovery pipeline from years of synthesis to computational screening — a guide to GNoME, the open databases, and what these tools still can't do.
Why this field is a strong proof point for AI in science
Materials discovery has traditionally moved at the pace of physical synthesis and testing — slow, expensive, and limited in how much of “chemical space” a lab can realistically explore. AI models trained on structure-property relationships have started to compress that search dramatically, generating and screening candidate materials computationally before anyone touches a furnace or a glovebox.
Core tools & databases
GNoME (Graph Networks for Materials Exploration)
What it does: A DeepMind “active learning” model that iteratively predicts new stable crystal structures, then folds its own predictions back into training data to improve further rounds. It identified 2.2 million new compounds — a dramatic increase over the roughly 48,000 stable materials humanity had discovered over decades of research — including 52,000 potential layered materials similar to graphene and 528 promising new lithium-ion conductors relevant to battery technology.
Access: Predictions are available through the associated Materials Project database; the underlying model itself is a DeepMind research release, not a self-serve product.
Best for: Screening for candidate stable compounds before committing to expensive synthesis and characterization.
Note: This is a discovery/screening tool, not a lab-automation tool — it tells you what might be stable, not how to make it.
Materials Project
What it does: Open, community-standard database of computed material properties (formation energies, band structures, elastic properties) for hundreds of thousands of known and predicted compounds, including GNoME’s predictions.
Access: Free, web-based, widely used as the default reference database in the field.
Best for: The natural starting point for any materials search — check here before assuming you need a new prediction.
Open Quantum Materials Database (OQMD) & AFLOW
What it does: Two other major open, DFT-based materials databases, each with somewhat different coverage and computational methodology from the Materials Project. Useful for cross-checking formation-energy predictions across independent sources.
Access: Free, web-based.
Best for: Validating a candidate material’s predicted stability against more than one computational source before investing synthesis effort.
Emerging: foundation models for battery materials (research-stage)
What it does: National labs and university groups are actively building broader “foundation models” that unify prediction of multiple material properties under one model, rather than training a separate model per property. Early examples target battery electrode and electrolyte materials specifically, with one such model being extended using exascale computing resources to a second foundation model for molecular crystals, the building blocks of battery electrodes.
Access: Not yet broadly released — research groups intend to make these available to the wider research community in the future.
Best for: Nothing yet, practically speaking — but worth tracking if you work in battery materials, since this is where the field’s leading edge is currently moving.
Suggested workflow
| Step | Tool | Purpose |
|---|---|---|
| 1 | Materials Project | Check if a candidate material or its properties are already computed |
| 2 | GNoME predictions (via Materials Project) | Look for AI-predicted stable candidates matching your target properties |
| 3 | OQMD / AFLOW | Cross-validate stability predictions against an independent database |
| 4 | Lab synthesis & characterization | Confirm real-world stability and properties — computational stability is not a guarantee |
What these tools can’t do yet
- Predicted stability isn’t synthesizability. A formation-energy calculation suggesting a compound is stable doesn’t mean it’s easy — or even currently possible — to synthesize in a lab.
- Databases can disagree. Different computational methodologies (Materials Project vs. OQMD vs. AFLOW) can produce different stability estimates for the same compound — cross-checking matters more here than in fields with a single dominant database.
- Property prediction still lags structure prediction. Predicting that a stable structure exists is more mature than reliably predicting how it will actually perform in a device (e.g., real-world battery cycle life) — treat performance predictions with more caution than structure predictions.
Related content
- Glossary: “Formation energy,” “DFT (density functional theory),” “Active learning”
- Field Guide: Structural Biology — for a comparison of how AI-driven discovery plays out in an adjacent field
- Field Guide: Climate & Earth Science