How Much Do Materials Databases Disagree? Materials Project vs JARVIS vs AFLOW vs OQMD

Each dot is one composition both databases have computed, placed at the band gap each reports for its lowest-energy structure. Points on the dashed diagonal agree exactly; the shaded band marks ±0.5 eV. Dots along an axis are compounds one database calls metallic (gap ≤ 0.05 eV) and the other does not.

All four databases run density functional theory, but they don’t make the same choices. Materials Project uses PBE and PBE+U, with r2SCAN for many entries in recent releases. JARVIS-DFT uses the OptB88vdW functional without a Hubbard U. AFLOW and OQMD use PBE, with U values for selected transition-metal compounds. A Hubbard U opens gaps in correlated d-electron systems. That is why NiO and LiFePO₄ look so different from one database to the next: 75% of the gap disagreements larger than 1 eV between Materials Project and JARVIS are oxides or fluorides of the eight elements Materials Project treats with a U.

Hull energies add a second layer. A compound’s energy above the hull is measured against every competing phase the database has computed in that chemical system. If one database has computed a more stable competitor, or never computed the experimentally observed structure, the verdict changes even when the underlying total energies are close. Each database also applies its own energy corrections to anions and elemental references.

None of this makes one database wrong. It means any single DFT number carries an uncertainty that the number itself doesn’t show. Our companion study, how accurate are DFT band gaps, measures that gap against experiment.

Same comparison for energy above the hull. Values are capped at 1 eV/atom for display, and the shaded band marks ±50 meV/atom. A compound is counted as on the hull at ≤ 1 meV/atom.

Ground-state band gaps (eV) and energies above hull (eV/atom). A dash means that database has no entry, or (for AFLOW) no band gap on its lowest-energy structure.

Compounds present in all three band-gap databases and in a measured band-gap compilation, ranked by the spread between databases. Most contain a transition metal or a rare earth, where the treatment of localised d and f electrons varies most between databases.

Measured compounds that sit on the hull in at least one database but 100 meV/atom or more above it in another. Gaps this large usually mean one database never computed the observed structure, not that the energies themselves disagree.

1. Every entry is keyed by its reduced composition (for example LiFePO₄ = FeLiO₄P), so differently written formulas from different databases line up.
2. For each database and composition, we take the lowest-energy entry as that database’s ground state. If that entry lacks a value, a sibling entry may stand in only when it is within 10 meV/atom of the ground state. This stops a gap from a high-energy polymorph being compared with another database’s ground state.
3. OQMD reports stability as a signed distance, negative for phases on the hull. We clamp it at zero to match the Materials Project and JARVIS convention. We drop <200 JARVIS rows (mostly elemental polymorphs) that report a zero hull energy alongside a large positive formation energy.
4. A gap ≤ 0.05 eV counts as metallic. An energy above hull ≤ 1 meV/atom counts as on the hull.
5. Charts show a random sample of up to 900 compounds per pair for legibility. Every statistic and the CSV use all shared compositions.
6. GNoME is excluded because its licence (CC BY-NC 4.0) does not permit commercial redistribution. Database snapshots: warehouse build 20260925T154847Z, page built 2026-09-30.