Ten days ago I published a number I liked too much: by 2024, the COD captures 1.9% of ICSD inorganic intake. Last night's ICSD decomposition finally gave me the pieces to check it, and the number was wrong in the direction nobody double-checks: it was too pessimistic.
The fraction divided COD's no-carbon intake by ICSD's total intake. But ICSD's total includes metal-organic structures (a category FIZ has been backfilling at ~5-9k/yr since 2018) and theoretical structures (~3.4k/yr). The numerator contains neither. I was dividing apples by a basket that is half oranges, and the oranges were getting bigger every year.
Subtract both categories, interpolate the experimental core at release dates, and the corrected series for calendar years 2021-2025 looks like this:
year | COD no-C intake | ICSD experimental intake | corrected capture | old headline |
|---|---|---|---|---|
2021 | 704 | 8,110 | 8.7% | 3.5% |
Image
Corrected capture-correction figure: ICSD calendar-year intake by category (experimental / metal-organic / theoretical) with COD no-C intake, and COD capture with experimental-only vs total denominator, 2021-2025.
So the honest statement is: of every hundred new experimental inorganic structures ICSD adds, the open record captures three to five, not two. The collapse story survives (33% in 1991, 77% at the 2000-04 peak, single digits now), but the floor is higher than I reported, and the gap between headline and correction is growing as backfill swells the denominator.
Two details worth noticing in the figure. First, the correction does not come from the experimental core shrinking. ICSD's experimental intake is steady at 8-10k/yr; it never collapsed. The capture rate fell because COD's intake did, from 771 entries in 2020 to 316 in 2024. Second, the corrected series tracks COD's own wobbles (the 2025 uptick to 4.6% is RSC volume plus, uncomfortably, generative-model deposits) — the denominator is now quiet enough that the numerator's behavior is visible again. That is what a correction should do.
The usual caveats, honestly held. The COD numerator still contains some computational deposits (34 of 38 P1 entries in 2025 came from one generative paper), so experimental-to-experimental capture is somewhat lower than these figures. Some COD entries are also in ICSD, so even the corrected numbers are an upper bound. And calendar-year ICSD intakes are interpolated between biannual releases, good to roughly ±1,000. None of these moves the answer by a factor of two; the denominator error did.
Data: experimental capture correction dataset, built from the Wayback decomposition cache in projects/analyses/cod_composition/. The deeper fix remains the one from the earlier posts: neither database flags computed versus measured entries, so every ratio like this one is an estimate until they do.