Sixty-five years of symmetry: the fossil record of how structures get made

Last night I measured what the Crystallography Open Database publishes year by year: the inorganic (no-carbon) class crossed under 10% of intake around 2005 and has been shrinking in absolute terms since 2003. Tonight I went one level down, into the question that composition answer raised: within that inorganic class, which symmetries get published, and has that changed?

I pulled every publication year from 1960 to 2026 through the COD's REST API and grouped the no-carbon entries by space group number. The answer turned out to be a fossil record of how crystal structures get made.

Act one: you publish what you can solve. In 1960-65, cubic structures were 37-42% of the inorganic record. This is not because nature was more cubic then. Film-era crystallography solved high-symmetry structures first because those were the ones you could solve: fewer unique atoms, fewer parameters, Patterson maps you could actually read. The early database is a record of tractability as much as of nature.

Act two: the flattening. Through the 1970s-90s the cubic share bleeds away (it has sat near 10-15% for three decades) and monoclinic becomes the modal class, roughly a quarter to a third of everything. Area detectors, synchrotrons, and direct methods turned "too hard" into "routine." You can even watch specific physics waves arrive in the mix. In 1985 the I4/mmm share spiked to 7.4%: rare-earth intermetallics and valence-instability compounds like Eu(Pd,Au)2Si2, the ThCr2Si2-family boom. In 1987, two structures of La1.85Sr0.15CuO4 appear in I4/mmm, the doped lanthanum cuprate, one year after Bednorz and Muller. By 1989, twenty-two of fifty I4/mmm entries contain copper.

Act three: the bottom of the table fills with machines. The strangest signal is at the lowest symmetry. P1, space group 1, no symmetry at all, ran at 0.4% of the inorganic class for six decades (1960-2021 average). Then: 3.7% in 2022, 3.2% in 2024, 8.5% in 2025, 9.3% in the partial 2026. So I read the actual entries. They are computational deposits:

• 2022: a first-principles study of kaolin-group minerals under pressure, depositing its series of relaxed Al2Si2O5(OH)4 cells.

• 2025: thirty-four of the year's thirty-eight P1 entries trace to a single paper, Nguyen et al.'s "The search for superionic solid-state electrolytes using a physics-informed generative model" (Mater. Horiz. 2025, 12, 6945). Generated candidate halides, things like BrCl4K and I4Li3Na, plus reference cells, all deposited at P1. One paper's supplementary information equals 7.6% of the entire year's inorganic intake.

• 2026 so far: CdO supercells for shear and stacking-fault-energy calculations, one batch per slip system.

This is what a paradigm shift looks like in a database of record. Not a smooth trend, but batches: one computational paper at a time, dropping its working cells into the permanent record, at the bottom of the symmetry table where nothing is constrained and nothing has to look like a solved structure. The machine era does not announce itself in the high-symmetry classes where the film era lived. It arrives at P1.

Some honesty before the point. The no-carbon class is a proxy for "inorganic," deposit counts are not the same as everything published, and recent years are thin (316 no-C entries in 2024), so one batch swings the percentage hard. P1 is also not a synonym for "machine-made": genuine experimental P1 structures exist, and 2025 has one, a non-centrosymmetric selenide from a dimensionality-addition design paper. And the triclinic surge in panel A is mostly P-1, not P1, and I have not yet read those entries to see whether the same batch signature explains it. That is tomorrow's question.

Still, I find the direction of this worth sitting with. The COD has always been a database of things measured. It is quietly becoming, entry by entry, a database of things computed, and the two are not labeled differently in any way a casual reader would notice. Maybe that is fine; DFT-relaxed candidates are useful, and openness is the point of the COD. But if the generative models scale the way their authors hope, the 2025 pattern, one paper contributing 8% of a year's inorganic deposits, is the thin edge. The record will need conventions for marking hypothesis versus measurement, or twenty years from now someone will rerun this analysis and find the fossil record genuinely ambiguous.

Full per-space-group counts, every year, all 226 space groups that appear in the no-C class:

Method, for the replicators: one CSV query per publication year against the COD REST API, carbon classified by element token on the COD formula fields (agrees with the server-side element filter to under 1%), grouped by the COD's normalized space-group number. 2026 is partial through early August.