One in six: how often open DFT databases disagree on whether a compound is stable
Most materials screening starts from one number in one database: an energy above the convex hull that says whether a compound should be stable, or a band gap that says whether it conducts. We wanted to know how much that number depends on which database you happened to open.
So we matched 62,803 compositions that appear in at least two of four open DFT databases (Materials Project, JARVIS-DFT, AFLOW and OQMD) and compared each database's ground-state answer. Most of the time they agree closely. The disagreements are concentrated, though, and they land exactly where screening decisions get made.
Stability verdicts are fragile
For every pair of databases we compared, 17–18% of shared compositions are on the hull in one database and above it in the other. That is roughly one compound in six.
The typical energy difference is small: the median hull-energy difference between Materials Project and OQMD is 3 meV/atom. The problem is that the verdict sits on a threshold. In 64–76% of the conflicts, both databases put the compound within 50 meV/atom of the hull, so a small shift in energy flips 'stable' to 'unstable'. And the tail is long: one in ten shared compounds differs by more than 232 meV/atom between Materials Project and OQMD.
Band gaps: metal or semiconductor depends on the database
Across the three band-gap pairs, 7–9% of shared compounds are called a metal by one database and a semiconductor by another, and 9–16% have band gaps that differ by more than 0.5 eV.
One modelling choice explains most of the large gaps. 75% of Materials Project vs JARVIS-DFT gaps that differ by more than 1 eV are oxides or fluorides of Co, Cr, Fe, Mn, Mo, Ni, V or W. Materials Project applies a Hubbard U to exactly these compounds; JARVIS-DFT applies none. They make up only 11% of the compositions the two databases share. LiFePO4 is 3.92 eV in Materials Project and 0.38 eV in JARVIS-DFT; NiO is 2.30 eV against 0.00 eV.
What this means if you screen materials
None of this makes one database wrong. Each makes different choices of functional, Hubbard U, magnetic initialisation and which competing phases it has computed. It does mean that a single DFT number carries an uncertainty the number itself does not show.
Two practical rules follow. First, treat any stability verdict within about 50 meV/atom of the hull as uncertain until a second source agrees. Second, for transition-metal oxides and fluorides, check which database a band gap came from before you rank on it.
This is why Lattice Graph shows every source side by side rather than averaging them. The full comparison, with charts, per-pair tables, the benchmark compounds and the outliers, is on the study page, and every row behind it is downloadable as a CSV with each database's identifier.