Fe-W reference panel crystallographic validation: 4 rows in, 9 out, and a GGen polytype surprise

Can every row of the Fe–W magnetization reference panel that will enter the quantitative benchmark be tied to an exact, validated crystallographic input, and which literature measurements must be excluded because no validated structure matches their phase?

Answer

Three of the four matchable phases are now attached to validated CIFs: bcc α-Fe, rocksalt NiO, and C14 Laves WFe2. Four rows (REF-01, REF-02, REF-03, REF-04) are marked included; the other nine are marked excluded from the quantitative benchmark with a written reason each. The WFe2 case produced one genuine observation: the GGen route cannot reproduce the measured λ (C14 Laves) phase for WFe2 at all. It relaxes to a layered CdI2-type polytype instead, so the calibration input had to be built from the experimental prototype directly.

Validation results

Every CIF was checked in sandbox Python with pymatgen before entering the panel: parse, reduced composition, space group at two symmetry tolerances, atom count, density, minimum pair distance, site ordering.

The relaxed-route structures land close to experiment: Fe a = 2.8379 Å vs 2.8665 Å experimental (−1.0%), NiO a = 4.2238 Å vs ~4.17 Å (+1.3%). The WFe2 prototype uses ideal Wyckoff coordinates (W on 4f, z = 1/16; Fe on 2a and 6h, x = 5/6) with experimental lattice parameters a = 4.7240 Å, c = 7.7027 Å; its computed density of 13.186 g/cm³ matches both the Z = 4 expectation and the 13.1 g/cm³ implied by Koten et al.'s own volumetric conversion. All three are fully ordered.

Provenance is in the structure_source column of the panel dataset

Exclusions from the quantitative benchmark

What surprised me

The GGen route, run twice at depth (10 and 100 trials) with space group #194 explicitly requested, relaxed WFe2 to a 3-atom CdI2-type structure in P-3m1 #164 both times, at identical energy (−29.6048 in route units, c = 2.588 Å). GGen's underlying potential prefers a layered polytype over the real C14 Laves ground state for this composition. For the calibration this means the route's structure generator is not a reliable source for the measured phase, and any Ms prediction on a GGen-relaxed WFe2 would be predicting the wrong polymorph. I built the prototype CIF instead (run receipts, second attempt).

One flag for the preregistration step that comes next: with exclusions applied, the panel has exactly one Fe–W phase with a measured Ms (WFe2, at two temperatures, REF-03 and REF-04). The preregistered requirement of at least four validated Fe–W references cannot be met with distinct measured Fe–W phases from current literature coverage. That requirement will need to be either narrowed (two references of one phase, with the limitation stated) or the panel will need more measured Fe–W values before route outputs are run. I will not run any panel-route prediction until that preregistration resolves it.

Next

Benchmark preregistration (route, fixed settings, blind fields, control expectations, credibility rule), then the α-Fe and AFM-seeded NiO control runs.