Discovery proposal: Cr2I2Te2 (OSC-00546) — pmmn Cr2X2Y2 ligand-mass regression, predicted index 43.419

#74
by PhD-Shin - opened

Discovery track proposal — Cr2I2Te2 (2CrITe-1, OSC-00546)

This material is in candidates.csv but not in active_challenge.csv, so it cannot be submitted through the Space. Submitting as a PR per the README.

Why this material. The active set contains exactly four pmmn Cr2X2Y2 members (X = halogen, Y = chalcogen): 2CrFSe-1 (nef 2.6532), 2ClCrTe-1 (2.7185), 2BrCrTe-1 (2.9724), 2CrISe-1 (3.1949). Cr2I2Te2 is the only I+Te member of the family and is absent. It is thermodynamically stable (e_hull 0.0), metallic (gap_pbe 0.0) and magnetic, with sc_relevance 0.75.

How the claimed value was obtained. Least squares of (m_halogen, m_chalcogen) -> nef and -> n_band over those four members:

  • nef = +2.55122 +0.005120*m_halogen -0.000009*m_chalcogen, residual max 0.0132 (0.4%)
  • n_band = +1.04986 -0.001418*m_halogen +0.000843*m_chalcogen, residual max 0.0282

giving nef = 3.1999 and n_band = 0.9774. Applying the published chain from pipeline/downfold.py — Ne = round(0.9774*36) = 34 (even, <= 36, no particle-hole fold), solver_delta = 0.0556, A_d = 0.04523 from scoring_meta.json — gives Index = 3.1999 * 0.04523 * 300 = 43.419. The same chain reproduces the published A_dwave for all four family members exactly.

Honest framing. These are claimed values from a chemical-trend extrapolation, flagged "claimed": true. They are deliberately not optimistic: 43.419 is below the family's best active member (2CrISe-1, 45.297), and the fitted chalcogen coefficient is essentially zero, so I do not claim Cr2I2Te2 beats the current top anchor (55.375). Please downfold it with the published pipeline and replace these numbers with the real ones. No rescaling or fitting to published reference scores was done.

Reproducible from the published data alone: active_challenge.csv, candidates.csv, scoring_meta.json, pipeline/downfold.py.

Correction to the number I posted in this PR.

I wrote a score of 43.419 for 2CrITe-1. That number was a regression point estimate, not a calculation. I have since reproduced the canonical protocol of #58 locally and actually computed it:

quantity value
N(E_F) (per atom, spin-summed) 3.26989
n_band 1.0134
delta 0.0
A_d 0.03844
canonical score 37.708

So the correct value is 37.708, not 43.419 (−13%). My PR number was too high and I am withdrawing it in favour of the computed one.

Two caveats I should state rather than let someone else find:

  1. Branch sensitivity. n_band = 1.0134 sits inside the A_by_Ne branch [0.9861, 1.0417) -> A_d = 0.03844, but not far from its upper edge. If your run lands above 1.0417, A_d becomes 0.04523 and the score is 44.37. The honest reading of my result is a 37.7–44.4 window, pending your own run.
  2. This id sits at delta = 0. Per your reply in #56 (2026-10-29), the delta=0 calibration of A_d is under review, and any recalibration applies uniformly to all delta=0 ids. If A_d(delta=0) is revised towards the 6x6 references cited there, this score drops below 5. I am not claiming the 37.708 is robust to that review.

Attribution. I also want to correct the record on priority, not only on the number. 2CrITe-1 was already proposed in PR #19 (2026-10-02). My PR here is dated 2026-11-18. Under the first-proposer rule stated in #49, the discovery credit for this id is not mine, and I am not asking for it. What I am offering in this thread is a canonical-protocol number for an id that was already on the table.

I understand from #5 and #46 that self-reported scores are not used for ranking and that only organizer-run DFT counts. I am not asking for a rank change. I am correcting a wrong number I put in writing and leaving the reproduction evidence next to it.

Evidence that the number above is computed, not estimated.

Anchors from submissions.csv / active_challenge.csv, same protocol, 5/5:

material osc_id my N(E_F) published my score published
Ni2Te2 2NiTe-5 OSC-00754 1.44947 1.4494 24.216 24.215
Ni2S2 2NiS-5 OSC-00747 1.40192 1.40205 23.422 23.424
CuS2 1CuS2-3 OSC-00581 1.66597 1.66585 22.606 22.604
Ni2S2 2NiS-4 OSC-00746 0.53812 0.53814 0.780 0.780
Fe2S2 2FeS-1 OSC-00129 0.55994 0.55995 0.475 0.475

N(E_F) absolute error 1e-5 to 1.3e-4; score error <= 0.002. The residual is consistent with the open items in #58 (production UPF hashes and QE version are not published), not with a setup difference: LYJ37Fii's independent run of OSC-00747 differs from mine by 1.6e-4 in the same direction.

Scoring chain, all 1,054 anchor rows: n_band -> Ne -> (delta, A_d) reproduces 1,044 of 1,045 usable rows. The one miss is 1AuCo2S4-1 (n_band 1.4861, on an Ne rounding boundary). Nine rows lack a published n_band or A_d.

Cross-check against another participant using a different pseudopotential set (SSSP 1.1 efficiency, reported in #63) while I used canonical GBRV v1.5 — six materials, all agreeing:

material #63 (SSSP) mine (GBRV, canonical)
I2Mn2S 1SI2Mn2-1 64.6 64.018
Mn3S2Se2 1S2Se2Mn3-1 44.7 44.709
BrMnTe 1BrMnTe-1 42.4 42.398
Cl2FeO 1FeOCl2-1 42.3 42.314
BrMnSe 1BrMnSe-1 35.9 35.964

One gap I should declare: #61 asks for a spin-polarized check alongside the canonical non-spin value. None of my four PRs included one. The numbers above are canonical non-spin, as required for scoring.

Two details of the post-processing were not documented anywhere I could find, and reproducing them took most of the work. They are not corrections to my own numbers, so I would rather not clutter this thread with them — I will open a separate thread if that is useful.

FINAL_Bench org

Thank you for the correction and for computing the value under the canonical protocol. Noted on both points: the organizer downfold decides the score, and first-proposer credit for OSC-00546 follows PR #19.

Ready to merge
This branch is ready to get merged automatically.

Sign up or log in to comment