Structures from a formula: symmetry-based generation through PyXtal

[Starlitnightly]

mv.gen.from_symmetry places the atoms of a composition on Wyckoff positions of a chosen space group and randomises the free parameters that remain.

With mv.gen.compositions from #12 the funnel now runs end to end — elements → compositions → structures → energies — where before it could only start from a structure somebody already had.

The output is a starting geometry, and says so

Cell volume is estimated and the free coordinates are random, so generated BaTiO₃ in Pm-3m comes out near 5.06 Å against a measured 4.00. That is the dangerous part of a structure generator, so it is stated in the registry notes, in uns["from_symmetry"]["note"], and in the tutorial.

Two columns, and the second is the one that matters

requested_space_group is what PyXtal was asked for. space_group is what pymatgen finds in the structure that came back.

They disagree far more often than I expected. Asking for 40 random groups for TiO₂:

attempts	40
structures built	7
failed (group cannot host the composition)	33
came back at a different symmetry	5 of 7

P-6 asked for, P6/mmm delivered — with few atoms the random free parameters land on a special position. A generator reporting only the request would be reporting its input. A test asserts the disagreement is real and that it is always upward.

The 33 failures are honest rather than hidden: a failure is a missing row plus a counted reason in uns["from_symmetry"], never a substituted structure. The notes say plainly that passing space_groups= explicitly is both faster and more likely to give what you asked for.

The probe caught another of mine

I declared obs["formula"] as required. The probe deleted it, the call succeeded anyway — the composition matrix is a perfectly good fallback — so the claim was wrong and is gone.

That is the third false claim of mine this contract probe has caught in this run of work, after a uns slot that landed on a different object (#11) and a plotting function that could not in fact plot what its notes promised (#9).

Verification

• 948 pass on py3.12; 846 pass on py3.10 where PyXtal is absent and its tests skip.
• Contract-verified rate stays 165/165. All 16 notebooks execute. Docs build.
• Tutorial output is generated, not described — the 7/33/5 numbers above are printed by the notebook itself.

🤖 Generated with Claude Code

https://claude.ai/code/session_016wh5NzNE9Wn49nuUm7zg2a