Real DFT as a level of theory, four more potentials, and two silent unit errors

[Starlitnightly]

matverse could reach six levels and every one of them was a surrogate. GPAW is now registered as gpaw-pbe / gpaw-pbe-fast with kind="dft" and surrogate=False, so the one level that solves the Kohn-Sham equations is distinguishable from the ones that reproduce something that did. m3gnet, m3gnet-r2scan, tensornet and orb are added alongside.

Validated by running it, not by citing it

Through matverse, on real PBE:

	matverse + GPAW	PBE literature
Si lattice constant (mv.calc.relax)	5.479 Å	5.47
Si bulk modulus (mv.prop.eos)	88.7 GPa	88–89
fit residual	0.002 meV/atom	—

Two things that would have been wrong silently

matgl returns stress in GPa; ASE's contract is eV/ų. Leaving stress_unit alone makes every stress-derived quantity — elastic constants, bulk modulus, pressure — exactly 160.2× too large, with no error raised anywhere. That factor is the GPa→eV/ų conversion, which is what makes it recognisable once you look. Fixed, with a test that fails on the unfixed version.

The second took longer and is not GPAW's fault. I first gave the DFT levels a k-point density — sensible for comparing different materials, wrong for scanning one material's volume. The mesh changes discretely as the cell grows, and each change puts a step in E(V) that the fit reads as curvature. Silicon, same calculator, only the k-points varying:

k-points	bulk modulus	mesh across the scan
density 2.0	−879 GPa	125 → 64
density 2.5	319 GPa	216 → 125
density 3.0	125 GPa	343 → 216
density 4.0	85.7 GPa	729 → 512
fixed 8×8×8	88.7 GPa	unchanged

A negative bulk modulus is not a soft crystal. Raising the density only shrinks the jump; it never removes it. The plane-wave cutoff was converged the whole time — PW(400), PW(600) and PW(800) agree to 0.1 GPa — so the entire error was the mesh.

obs["eos_residual"] catches it long before the modulus does: 9.5 meV/atom against 0.002, a factor of nearly five thousand. The advice already in that function's notes — a residual far above a meV is a fit that should not be read — was enough to reject it without knowing the cause.

pymatviz: a judgement, not a wholesale add

Most of it duplicates mv.pl, which already has hull, parity, periodic_table, spectra, scatter, bands and distribution. Wrapping those again is the mistake the coverage map exists to prevent. The one genuinely absent capability is the space-group distribution, so mv.pl.spacegroups does that directly — grouped and coloured by crystal system rather than plotted as 230 flat categories, and able to overlay requested against achieved for mv.gen.from_symmetry output.

mv.pp.symmetry gained spacegroup_number and spacegroup_symbol while I was there: it was already building the analyser, and a symmetry report giving the crystal system and point group but not the group itself is missing the one label everybody cites.

Verification

• 960 pass on py3.12 with GPAW and matgl; 850 on py3.10 without them.
• The level-metadata tests skip where a backend is absent rather than failing — which they did not do at first, and py3.10 caught it.
• Contract-verified rate stays 165/165. All 16 notebooks execute. Docs build.
• CI gains dft: the GPAW test is one H₂ molecule in LCAO, ~3 s. The crystal validations behind the numbers above take minutes each and stay out.

🤖 Generated with Claude Code

https://claude.ai/code/session_016wh5NzNE9Wn49nuUm7zg2a