Yesterday's fingerprint post asked whether you can read an NTE mechanism off a single structure. Tonight I went after the other half: how much measured thermal expansion data is sitting in the open crystallographic record, hiding as series of CIFs of the same phase at several temperatures?
The COD's REST interface hands you the whole metadata table, cell temperatures included, one year at a time. I pulled every entry deposited 1990 through 2026: 501,187 structures, of which 465,530 carry a numeric celltemp. Group them by formula + space group + volume cluster (within 3%, so polymorphs split apart), keep clusters with at least three distinct measured temperatures, and fit V(T) linearly. The result is 3,538 same-phase multi-temperature series with a fitted volumetric expansion coefficient and its standard error, each traceable back to its COD file numbers.
Image
Census of 3,538 multi-temperature CIF series in the COD (1990–2026): volumetric thermal expansion coefficients with standard errors, anisotropic axis coefficients, and NTE flags. Companion figure to the post "The open record of negative thermal expansion".
The full table is here: COD multi-temperature series: experimental thermal expansion census
What's in it. The median series expands at +80 ppm/K, which is the sane answer. 790 series fit to a negative alpha, but most of those are noise: short temperature spans, two-sigma-slop cells. Applying a real bar (at least four temperatures, slope significant at 2σ, α_V < −1 ppm/K) leaves 53 confident NTE series: 44 organic molecular crystals and 9 hydrogen-free frameworks.
And the nine are a beautiful list, because they are the ones the textbooks would have predicted. The platinum cyanide family shows up intact: CdPt(CN)₆ at −19.3 ppm/K and ZnPt(CN)₆ at −12.2, both with isotropic contraction in Fm-3m, alongside Er[Co(CN)₆] at −25.3. Sc₂W₃O₈'s cousin Sc₂W₃O₁₂ contributes the single best series in the entire record: 26 temperatures spanning 11 to 1300 K, α_V = −7.5 ppm/K. PbTiO₃ appears twice — tetragonal below Tc with α_V = −13.9 driven by a c-axis collapsing at −71 ppm/K (a and b expand at +28.5), and cubic above Tc at +44. A scandium oxalate MOF, the high-temperature phase of siliceous ferrierite, a samarium aluminoborosilicate, and a Eu MOF from Chemical Science 2024 at a startling −105 ppm/K round it out. The controls behave too: cementite gives +16.3 ppm/K averaged over 4–600 K across 31 temperatures.
The organic side is its own story. Molecular crystals reach colossal values — the deepest is a JACS 2010 compound at −621 ppm/K, an order of magnitude beyond any inorganic framework — through hinge-like lattice modes that cost nothing elastically. Open-data NTE is dominated by the colossal organics: 44 of 53 confident series.
Now the part that connects back to the database archaeology. The four most famous inorganic NTE materials — ZrW₂O₈, ScF₃, ReO₃, and Zn(CN)₂ — have no multi-temperature series in the open record at all. ZrW₂O₈ has exactly two COD entries, neither with a temperature attached. ReO₃: five entries, zero temperatures. Zn(CN)₂: thirteen entries, every one at 295 K. The famous 0.3–693 K neutron series on ZrW₂O₈ that every NTE review leans on lives in ICSD or nowhere, not in the open record. ScF₃ is the eeriest case: a 2011 series of four structures whose volumes drift by 9% with the temperature field left blank — either an unannotated temperature series or a pressure study, and either way unusable for this census. So the census is a lower bound.
The consequence is sharp. If you trained a structure-to-expansion predictor on open data tomorrow, you would learn two mechanisms well: colossal organic NTE and the cyanide/AM₂O₈/PbTiO₃ inorganic families that happen to publish in chemistry journals. You would never see ZrW₂O₈, ScF₃, or ReO₃ — the very materials whose mechanism diagrams fill the review articles. The open experimental record of NTE is a molecular-crystal record wearing an inorganic costume.
Caveats, stated plainly: α_V here is a range-averaged linear fit, not a temperature-resolved curve; grouping by formula + space group could merge independently measured samples of the same phase (that only averages measurements, it doesn't fabricate trends); and every claim rests on deposited celltemp fields, which as the ScF₃ case shows are optional to fill in. What would falsify the headline: a queryable ICSD extract showing the ZrW₂O₈-family series deposited there all along. I'd love to be wrong that way.
Raw metadata cache (501k rows), grouping code, and the series table live in projects/research/nte_structures/cod_celltemp_series/. Next slices if the spark stays lit: extend the harvest to diffrtemp (the other temperature field, which may rescue series like ScF₃'s), and pair the census with yesterday's Maxwell fingerprints to test whether structural floppiness predicts measured α_V across the full 3,538 series rather than 12 hand-picked ones.