BIGGER PICTURE Scientific discovery
Millions of imagined crystals, one real battery
Microsoft’s models shortlisted a solid electrolyte; PNNL chemists made it and built a functioning battery. The surprise is the bridge from virtual screening to a physical object—not a market-ready battery breakthrough.

- Initial virtual candidates
- 32.6 million
- Expert shortlist
- 18 materials
- Candidates to working cell
- Under nine months
How the idea works
Generate crystal variants
Swap elements in known structures and predict useful properties.
Apply stricter physical tests
Simulations and experts reduce millions of candidates to a shortlist.
Make and measure the material
Synthesize the electrolyte and test a real battery cell.
Search the crystal possibilities
The search began by substituting elements in known crystal structures, generating 32.6 million candidates. Making each one would be impractical. AI models first estimated properties such as stability and electrical behaviour, reserving slower physics calculations for the survivors. The target was a solid electrolyte: the material through which ions move inside a battery.
Sources: [2]
Let physics challenge the prediction
The pipeline did not treat a model score as experimental truth. More accurate calculations checked the predicted properties; molecular simulations explored how the atoms and ions would move. Researchers then considered practical constraints, including available elements, to select 18 candidates. This sequence combines broad machine search with increasingly expensive tests and scientific judgement.
A powder had to become a battery
At PNNL, chemists synthesized the selected compound, examined its structure and measured conductivity. Its chemistry combined sodium and lithium. Crucially, the team built and tested an all-solid-state battery rather than stopping at a promising computer prediction. PNNL reported less than nine months from receiving simulated candidates to producing a functioning cell.
A faster experiment, not a finished product
PNNL explicitly said that additional optimization was needed to compete with published benchmarks. The demonstration does not establish a longer-lived electric-car battery or a commercial production process. Its practical contribution is a research workflow: expend laboratory effort on a much smaller, evidence-informed set, then use real measurements to decide whether the material deserves further development.
Sources: [1]