Rare-earth and platinum-group metals are over-represented in AI-predicted crystals
Google DeepMind's GNoME project predicted 554,054 stable inorganic crystals. This audit asks which elements those crystals need. Compared fairly with known materials, they contain rare-earth and platinum-group metals far more often, and these are elements whose supply depends on very few countries.
How to use this page
GNoME materials contain more critical elements than known materials
Each group is compared within the same compound types (same number of elements, metal-only or with an anion), so differences in the kind of compound do not count as a difference in elements.
__HINT__Choose a reference group to change the chart. Hover or tap a bar for the exact numbers.
Heavy lanthanides and platinum-group metals are the most over-used elements
The colour shows how much more often GNoME materials contain each element than known materials do. Violet means over-used, teal means under-used. The size of each ratio depends on how the known materials are matched, so both versions are shown.
__HINT__Click any element to see its numbers. Switch the matching to compare the two methods.
Dashed cells: not enough compounds to compare. A small dot marks radioactive elements, whose very large ratios rest on few known compounds.
Much of the gap comes from the elements chosen, not only the kind of compound
The gap can arise in two ways: GNoME may have more of the compound types that often contain these elements (type mix), or it may choose these elements more often inside the same type (within type). The split depends on whose type weights are used.
__HINT__Switch the weights to see how the split changes.
Share of the gap on the common support. Averaged over both weight orders, more than half of each family's gap arises within the same compound type; for platinum-group metals the answer depends on the order.
Compared with random element choice
1.0 means a group uses the family as often as a random choice of elements would. Known materials sit well below 1, GNoME above it.
The gap has two sides. Known materials reflect what chemists have chosen to make, which may be shaped by cost and availability. GNoME also lies above the random baseline and above other computed databases for rare earths, so the dataset itself adds to the gap.
Rare-earth materials lie closer to the stability limit
A material is near-hull when its predicted energy is less than 25 meV/atom below the convex hull of competing phases. Materials with lanthanides whose 4f electrons are frozen in the calculation are near-hull more often.
__HINT__Switch the reference hull. Hover a bar for the numbers.
Odds of being near-hull with the family, matched by compound type. The first interval treats every compound as independent. The second interval resamples near-copies (compounds that differ only by swapping one rare earth for another) together, so it is wider and more honest.
The excess reaches application shortlists, at a smaller size
Energy-GNoME shortlisted GNoME candidates for battery cathodes, solar perovskites and thermoelectrics. Their critical-element shares are compared with known materials, and each high-risk candidate (EU atom share of 0.5 or more) is checked for a lower-risk alternative with similar predicted performance.
__HINT__Pick a shortlist and a matching method. The table below responds to the rule you choose.
__HINT__Choose a rule for the table below. Stricter rules leave fewer alternatives.
An alternative has predicted performance within 10% (0.1 eV for perovskites) and an EU atom share at least 0.2 lower. Performance values are machine-learning predictions, so alternatives are leads for study, not recommendations.
Supply depends more on a few countries
Each element is linked to its world mine production (USGS 2025). The bars show what share of the atoms in each group come from each producing country, with known materials matched to GNoME's compound types.
__HINT__Hover or tap a bar for the exact shares.
Rare earths, four platinum-group metals and hafnium share one production pattern per group in the USGS data, so these figures are approximate. They describe mining, not refining.
The main result holds under the checks a reviewer would ask for
These checks were added after an internal review. Each one changes one choice in the analysis and asks whether the rare-earth gap survives.
Full tables are in the electronic supplementary information of the paper and in the files R1 to R8 of the results folder.
What this study cannot show
- It looks at the elements materials contain. It does not show whether any material can be made.
- Known materials reflect the choices of experimental chemists, so the gap is measured against experimental practice, which has its own preferences.
- Many GNoME entries may be ordered versions of disordered phases. The study does not correct for this.
- It does not test the frozen-4f approximation. That needs calculations with explicit 4f electrons.
- Five-element metal-only compounds can only be compared with other computed databases, and the four-element metal-only type rests on 276 known compounds (the result holds without it).
- Application performance values are machine-learning predictions.
- Country data come from mine production and group some elements together.
- All results refer to the August 2024 release of the GNoME data, and the audit has so far been applied to this one dataset.