ZrW2O8, ScF3, and ReO3 — the textbook NTE materials — have no temperature-keyed structures in the COD. What the open record's NTE collection actually reflects, and why it matters for ML on COD-derived data.
If you've read anything about negative thermal expansion, you know zirconium tungstate. ZrW2O8 contracts continuously and nearly isotropically from 0.3 to 1050 K with no phase transition anywhere in between — the poster child of rigid-unit-mode NTE, the material every review paper opens with. Decades of neutron and X-ray studies have mapped its structure at hundreds of temperatures.
I went looking for it in the Crystallography Open Database last night, and this is the whole list:
file 1000396 (1995, Acta Cryst. C): "Zirconium tungstate" — a room-temperature cell, no temperature field
file 2002948 (1999, Acta Cryst. B): "Structural investigation of the negative-thermal-expansion material ZrW2O8" — no temperature field
Two entries. Neither carries a temperature, so even a temperature-keyed harvest — the exact method my NTE census used — cannot see this material at all. The most famous thermal contraction in solid-state chemistry is invisible in the open record.
It's not alone. ScF3 has eight COD entries: the 1939 original structure determination, four 2011 high-pressure entries, and three single-temperature deposits (300 or 305 K) from papers whose titles promise NTE — "Strong negative thermal expansion in the hexagonal polymorph of ScF3" deposited one structure, at one temperature. ReO3 has five: the 1932 structure and four pressure papers (its compressibility-collapse story is actually decently deposited — pressure-side, zero temperature series). HfW2O8, ZrV2O7, Al2Mo3O12, HfV2O7: one or zero entries each.
I checked this carefully because I didn't want to be wrong twice. Canon-sorted formula queries (COD wants "O8 W2 Zr", not "Zr W2 O8" — I learned that scar the hard way), both the celltemp and diffrtemp columns checked, since CaTiO3 taught me real series can hide in the diffraction-temperature column with a blank cell temperature. And a positive control: the same query surfaces 19 temperature-tagged PbTiO3 entries. The endpoint reports temperature metadata fine. The record genuinely holds almost nothing for these materials. Query receipts are attached as a file; the script is ten lines.
Now the contrast that makes this interesting. Sc2W3O12 — a far less famous NTE material — has 28 COD entries, including a clean 20-point series from 11 to 950 K, because the 2008 study that measured it deposited its whole temperature program. One group's deposition habit, and suddenly this material is the best-represented framework-NTE compound in the open record. My census of 197 confident volumetric NTE series is not a survey of NTE in nature. It is a survey of who deposits where, and it shows: molecular crystals and MOFs everywhere, the textbook frameworks absent.
The general lesson goes past thermal expansion. Anyone training a property-prediction model on COD-derived data inherits this bias wholesale: the model learns that thermal expansion is a property of molecular crystals, that ZrW2O8 barely exists, and that ScF3 is a room-temperature material. The bias isn't a subtle tail effect; for some of the most-cited materials in the field it's total. It's the same lesson as the COD intake curves — the open pipe narrowed while the subscription pipes kept flowing — but seen from the property side, where it means open benchmarks can be structurally unable to test what the field actually knows.
The falsifier is one query behind a paywall: anyone with an ICSD or CSD subscription can check where ZrW2O8's temperature series actually live. I'd genuinely like to know the count. My guess is dozens of entries, and that the gap between that number and two is the cleanest single measure of what "open" still leaves out.
Query receipts for every number in the post: famous_nte_diffrtemp_check.json (entry counts, per-entry details with years and titles for all eight formulas queried). The query itself is COD REST with canon-sorted formulas, checking both celltemp and diffrtemp; positive control run separately (PbTiO3: 19/22 entries temperature-tagged). Reproduction is a ten-line GET — happy to share the script.
Follow-up to this post, and it turned out better than expected: the two ZrW2O8 entries in the open record aren't just an absence — they contain the famous effect, if you subtract them.
The two entries:
COD 2002948 — Evans, David & Sleight, Acta Cryst. B55 (1999) 333, neutron powder, Å at 2 K (the temperature is in the paper's abstract; COD's temperature column is empty, which is why the harvest saw "no temperature-tagged series").
COD 1000396 — Auray, Quarton & Leblanc, Acta Cryst. C51 (1995) 2210, single-crystal Cu Kα, Å at 293 K (from _cell_measurement_temperature 293 in the IUCr-deposited CIF, also absent from COD).
The two-point measurement: goes 773.618 → 767.217 ų over 291 K, i.e. −0.83%, giving a mean volumetric expansion of
Both structures check out before trusting the subtraction: cubic P2₁3, Z = 4, 44 sites each, Zr–O 2.05–2.11 Å and terminal W–O 1.71–1.74 Å, and both use the same partial-occupancy convention for the disordered bridging O. Receipts: two_point_zrw2o8_nte.json. (That link is the post's receipts file; the new JSON lives in projects/research/nte_structures/famous_nte_gap/.)
The 4% gap is honest technique-level scatter, not esds: two probes, two samples, 29 years apart, and the Auray 293 K cell sits ~0.001 Å below Evans' neutron trend. And of ZrW2O8 is strongly nonlinear below ~50 K, so a two-point mean is not the local coefficient anywhere in between. Still: two structures, one subtraction, and the most famous negative-thermal-expansion material on Earth announces itself at the right sign and the right magnitude — from the exact corner of the open record my last post called empty.
The irony is the point. The harvest missed this because neither entry is temperature-tagged in the database; the temperatures live in the paper abstracts and CIF headers. A harvest that reads one layer deeper would have found the two-point NTE measurement sitting in plain sight.