Do the energy model's training labels get the molybdenum–tantalum mixing energy right?

EXPERIMENTS.md · lines 11481–11534

E180 — Mo₂₇Ta₂₇: the basin against the literature, and a test of the labels themselves

The external numbers. Blum & Zunger (PRB 69, 214202 / PRB 72, 020104; LDA mixed-basis cluster expansion): the fully random Mo–Ta at x = 0.5 has ΔH_mix = −127 meV/atom; the measured mixing enthalpy is −114 ± 26. Widom (arXiv:1306.5043, PBE): ordered MoTa (oA12/B2) −186 meV/atom, Mo–Ta the strongest-bonding pair of the MoNbTaW binaries.

The internal numbers. icet rung 0 for Mo–Ta 50/50: e_bcc −11.5 meV/atom, T_order 1782 K (above the window: a pass), p 0.997. eCE for the seed-0 Mo₂₇Ta₂₇ decoration: −23.6. And the finding that changes what this cell tests: RHEA's own 54-atom random Mo–Ta 50/50 cell carries a label of −22.9 meV/atom in my corrected label set (the eCE gives it −9.9; the three 16-atom Mo–Ta cells carry −133, −29, −21). The model is faithful to its labels. So either (a) the label pipeline — rattle correction by MACE difference, volume relaxation, strain cut, pure-cell referencing — has removed ~100 meV/atom of mixing energy from random Mo–Ta cells, or (b) RHEA's PBE-VASP random Mo–Ta really is at −23 and the LDA-CE literature is 5× away, which no code difference explains.

Cell. Mo₂₇Ta₂₇, seed-0 decoration on ideal sites, Vegard a = 3.2460 Å (L 9.7380), E176's settings, single SCF; references Mo and Ta from refs_v2/.

Predictions — two branches, one of which this cell falsifies.

• Branch A (literature): E_form(DFT, QE) in [−150, −100]. Then RHEA's raw VASP energy for its random cell must also be near −120 and my correction took it to −23: the label pipeline is the defect, every "compression" measured this week is largely a label artefact, and E172/E176/E177 must be re-read against raw labels.
• Branch B (labels): E_form(DFT, QE) in [−40, −5]. Then the labels are right, the QE cell agrees with RHEA's VASP, and the literature random value is not comparable (LDA, MBCE fit) — the cheap rungs are consistent with DFT at this composition and E177's −71 was the Ti/Hf/V admixture, not the Mo–Ta basin.
• Anything between −100 and −40 is neither and is stated as such. The raw-versus-corrected label for RHEA's Mo–Ta cell is read before DFT finishes and recorded beside this, so the DFT verdict lands on a stated expectation.

Read before DFT — and it is my pipeline. RHEA index 6369 (54-atom random Mo–Ta 50/50, bcc_alloys): raw E_DFT −11.3426 eV/atom, rattle+volume correction +0.0858 eV, E_lattice −11.4284. Referenced to corrected pure cells: e_formation = −73.7 meV/atom (rhea_labels_formation_strain0.10_pureref.jsonl). Referenced to the regression intercepts: −22.9 (rhea_labels_formation_strain0.10.jsonl). The deployed model, ece_v3b_strain010, was fitted on the second file — the training property for this cell is −22.9, and the eCE reproduces it. So:

• The regression references absorb most of the mean mixing energy into the "elements", and formation energies referenced to them are compressed toward zero by construction. That is the 0.59 holdout slope, and it is the 1.9× / 2.2× / 4.0× of E172, E176 and E177 — not the model extrapolating, as E176 and E177 concluded. Those readings are withdrawn in part: the model may still compress on unseen compositions, but that cannot be measured until it is trained on labels whose zero is the elements.
• The memory note "regression references ≠ elemental energies → use corrected pure cells" existed; --pure-references was built; the model that went into service was the one fitted before it. A rule for the record: the label file's basename is part of the model's name from now on.
• E180's DFT now lands on three stated expectations: literature −127 (LDA, random), the pure-referenced label −74 (PBE-VASP, RHEA's own random cell), the regression label −23. Same-code agreement with RHEA is the −74; the QE cell should sit within ±12 of it if the two PBE-PAW codes agree on Mo–Ta as ACWF says they should.

EXPERIMENTS.md · lines 11548–11558

E180 result — between the two branches, and on the pure-referenced side. JOB DONE in 21 iterations; E/atom −7422.32794 eV at Vegard; references Mo, Ta (refs_v2/): E_form(DFT, QE) = −89.8 meV/atom. Branch A's window [−150, −100] missed by 10; branch B's [−40, −5] missed by 50. Against the three stated expectations: literature random (LDA MBCE) −127: 37 off; RHEA's pure-referenced label −73.7: 16 off; the regression label −22.9: 67 off. Inside the measured mixing enthalpy −114 ± 26. So: the regression-reference label is falsified as a label; the pure-referenced convention is the right one; and the 16 meV between the QE ideal-site cell and RHEA's corrected cell is the size of the rattle correction itself (+86 meV on a 0.32 Å-rattled cell) — the correction's error, not a code offset (Mo and Ta are V-free, no cutoff caveat). The icet CE (−11.5) and eCE v3b (−23.6) were 4–8× off here, as E177 already showed for the basin.