Where did the structures go?

This is the fourth post in a series I did not plan. It started with a sanity check on random COD structures, which led to the composition flip of 1995, then to sixty-five years of space-group history, then to an entry-by-entry autopsy of the triclinic surge. Every one of those posts kept tripping over the same denominator: the Crystallography Open Database's inorganic (no-carbon) intake peaked at 2,949 entries in 2003 and fell to 316 in 2024, the lowest since 1961.

The question I kept dodging: where did the structures go? Did experimental inorganic crystallography quietly die?

No. It is bigger than ever. The paywalled Inorganic Crystal Structure Database held about 100,000 entries in 2007, passed 200,000 in 2018, and stands at 335,032 in its 2026.1 release, adding roughly 16,000 structures a year. The Cambridge Structural Database passed one million structures in June 2019 and reached 1.34 million by the end of 2024, growing at over 60,000 a year. The field is producing structures at record rates. They are just no longer arriving at the open database.

I pulled every no-carbon COD entry since 1990 (46,753 of them, via the REST API) and grouped each by the publisher family of its source journal.

The anatomy shows three exit waves, each publisher on its own schedule. Elsevier and Wiley cut within a year of each other: Elsevier went 902 entries in 2003 to 783 in 2004 to 145 in 2005; Wiley went 201 (2004) to 64 (2005) to 16 (2006). The mineralogy journals, the field's backbone at 400 to 600 entries a year for two decades, held a long plateau and then cut in 2014-15: 192 to 45 in a single year, 26 the next. ACS was the last giant standing; Inorganic Chemistry alone contributed 539 entries in 2013 and the ACS family peaked at 726. Then 647 (2014), 559 (2015), 136 (2016). By 2024: zero.

Who still feeds the open record? Essentially the RSC. Dalton Transactions, CrystEngComm, and their siblings contributed 315 of the 449 no-carbon entries in 2025, about 70%. The IUCr journals add a small guaranteed trickle because they require CIF deposition as a condition of publication. Beyond that: a long tail of small journals, and, increasingly, deposits with no journal at all, the personal communications and computational working directories behind the P1 and P-1 spikes. (One wrong guess worth preempting: the famous Acta Cryst. E structure mill is not the answer here. At its 5,181-papers peak in 2007 it was 61% organic and 36% metal-organic, only 3% inorganic.)

So the structures didn't vanish; the deposits changed destination. Publisher data workflows now route inorganic structures to ICSD, whose staff curate from 80 leading and 1,300 further journals behind a subscription, and anything containing carbon to the CCDC deposition-number system. The COD stopped being wired into the pipe.

Put as one ratio, measured against ICSD's annual intake: the COD's open capture of inorganic crystallography ran near 15% in the early 2010s (1,443 entries in 2013 against ICSD's then ~9,400 a year) and sits around 2% now (316 against ~16,800 in 2024). That ratio is an estimate, and the early-2000s peak was itself inflated by back-catalog digitization (two-thirds of 2003's entries carry no DOI), so read it as an order of magnitude, not a measurement.

This recolors the whole series for me. The composition flip, the decline of cubic, the triclinic spikes: these are stories about capture, not production. A database that loses its feeds becomes a funhouse mirror, where a shrinking denominator turns niche flows, a generative-model batch here, a directory of shear supercells there, into apparent trends. Anyone mining the COD for what chemistry happened now needs a capture correction. I certainly do.

Honest limits. The mechanism is inferred from timing: staggered, per-publisher cutoffs argue for publisher-side workflow changes rather than one COD-side failure, but I have no contracts. ICSD's intake includes back-curation of older literature, which flatters its rate slightly. The no-carbon class excludes organometallics by construction; those went to the CSD all along. And COD's method field is too sparsely populated (3,000 of 47,000 entries) to test whether powder and single-crystal work left at different speeds.

One closing thought. The COD is still the only place where you can run this entire analysis in an afternoon, for free, without asking anyone's permission. The RSC's constancy shows that a publisher can keep feeding the open record if it simply decides to. The open question for the next decade is whether anyone else decides it is worth wiring back in.

Series data: publisher-family counts