How far to trust a computed property

Last verified 19 August 2026.

Why an engineer should care about this at all

Because the numbers are already in front of you. Materials Project, OQMD, AFLOW, JARVIS and the rest are open, free, growing, and increasingly what a search or an AI assistant returns when you ask about a compound. We read them too.

They are not measurements. They are calculations, and calculations have characteristic errors that do not resemble measurement scatter — they are systematic, they differ by property, and they differ by which functional was used. An engineer who treats a computed number as approximately-measured will be wrong in a specific direction, repeatedly.

Band gaps: about 40% low, and sometimes qualitatively wrong

This is the best-documented case, and the Materials Project says it about its own data.

The cause is structural, not a tuning problem: approximations in the exchange–correlation functional, plus a derivative-discontinuity term that standard functionals do not capture and which is generally regarded as the larger contribution.

What remains reliable is the shape: band dispersion carries errors of roughly 0.1–0.4 eV, so the qualitative picture — where the bands go, which are flat, where the extrema sit — is useful even when the gap value is not.

Source: Materials Project documentation, electronic structure methodology.

Formation energies: roughly 100 meV per atom, depending on the functional

Kingsbury, Persson and co-workers benchmarked functionals against 986 materials with experimental data (Physical Review Materials 6, 013801, 2022).

r²SCAN gave roughly 20% lower MAE than SCAN for strongly bound materials and 15% for weakly bound ones. Volume errors came from a separate set of 4,974 materials; band gap errors from 582.

Two things an engineer should take from that table. The functional matters as much as the compound — quoting a formation energy without saying which functional produced it discards information that changes the number. And geometry is far better predicted than energy: lattice volumes land within about one cubic ångström per atom, which is why computed structures are genuinely dependable while computed energetics need more care.

Source: Kingsbury et al., Phys. Rev. Materials 6, 013801 (2022).

What this means on the evidence ladder

A DFT value is E2 — computed and transparent under the E-ladder, and only when the record carries the method, the code, the functional and a resolvable entry identifier. Strip the mp-id and the functional off it and it is not E2 any more; it is a number whose error bar nobody can reconstruct.

Note where E2 sits: below a datasheet value at E3. That ordering surprises computational scientists and it is deliberate. The ladder measures how far a value can be defended in design documentation, not scientific merit. A DFT result is often more reproducible than a datasheet figure and still not a measurement of the material you are going to buy.

The specification says this ordering is contested and lists it as an open question. If you think it is wrong, that is a useful argument to have in public.

Where computed data is genuinely strong for engineering work

Screening large spaces. Ranking hundreds of compositions for stability or for a target property is exactly what DFT databases exist for, and systematic error largely cancels within a ranking.

Structures and geometry. Lattice parameters, symmetry, coordination. The volume errors above are small enough for practical use.

Trends and comparisons. How a property moves with substitution, across a family, along a series. Differences are more reliable than absolutes.

Phase stability, with care. Convex-hull distances are the standard use, and they inherit the formation-energy error above — a compound sitting 30 meV/atom above the hull is inside the noise.

And where it is not

Design allowables. Never. No computed value is a specification minimum, and nothing in this page suggests otherwise.

Absolute values for a property you will test. If the number will be compared with a measurement, expect the offsets above.

Mechanical properties of engineering alloys. DFT operates on ordered crystalline cells. Yield strength in a real alloy is governed by microstructure, grain size, dislocations, precipitates and processing — none of which are in the cell. This is why a computed database has nothing to say about 316L.

Whether the compound can be made. A stable entry in a database is not a synthesis route, and an entry existing is not evidence anybody has made it.

The practical habit

When you take a number out of a computational database, record four things: the entry identifier, the functional, the property, and the date. Those four turn a bare figure into something you or a reviewer can re-derive later — and they take about ten seconds, which is the entire cost of the difference between E2 and E0.

And where a computed value is doing real work in a decision, say so in the record explicitly, alongside the error width from the tables above. A computed number carried into a design note without that context is the same failure as a typical value used as a minimum, wearing different clothes.

What this page does not cover

• How DFT works. Plenty of good introductions exist; this is about how far to trust the output.
• Hybrid functionals, GW and beyond. More accurate, far more expensive, and rarer in the open high-throughput databases most people encounter.
• Machine-learned interatomic potentials. MACE, CHGNet, M3GNet and their relatives are a different question, sit at E1, and deserve their own page.
• Error bars for a specific compound. The figures above are benchmark aggregates. Your compound may be better or worse and the aggregate does not tell you which.
• Corrections schemes. Various exist for formation energies and for gaps; this page does not assess them.