Derived data and analysis artifacts for: Computed materials proposals depart from the structural memory of experimental discovery

Production analysis artifacts accompanying the manuscript “Computed materials proposals depart from the structural memory of experimental discovery.” Author note: D. Nguyen and K. Cao are designated joint first authors and contributed equally to this work. Generative AI and high-throughput density-functional-theory pipelines now claim millions of new crystal structures, but the field lacks a calibrated reference frame for comparing such proposals with experimentally realized inorganic chemistry. The paper embeds 167,500 Inorganic Crystal Structure Database (ICSD) entries in a continuous structural-similarity space, partitions the map into Louvain graph communities, replays the partition through publication time, and projects five publicly-released external structure samples (GNoME, MatterGen, Materials Project theoretical-only, JARVIS-DFT off-hull, Alexandria-PBE off-hull) into the same frozen historical geometry under multiple held-out training cutoffs. This Zenodo deposit holds the artifacts that are too large or too binary for the paper’s GitHub source repository — the frozen 167,500-entry feature matrix, the methods-text-faithful ablation features, and a single tar.zst bundle of the per-claim run outputs. Code, manuscript sources, the lightweight CSV/JSON artifacts consumed by figure scripts, the dash app, and the student/reviewer code guide are version-controlled in the accompanying GitHub repository (URL to be linked at publication). Contents features.npy — frozen ICSD feature matrix, 167,500 × 213 float64. The structure-level descriptor produced by the production featurization pipeline before PCA. Row-aligned with the canonical Louvain partition CSV (in the tar bundle and in the GitHub repo’s notes/icsd_community_assignments/). features_pca.npy — the frozen 32-dimensional PCA basis, 167,500 × 32 float64 (76.8% of standardized variance retained). What every external-source projection scores against; load this to project a new structure into the canonical ICSD basis. features_pca_paper_text_ablation.npy — the PCA(32) basis under the methods-text-faithful alternative pipeline (matminer Magpie-22 + concat-mean-std Weisfeiler-Lehman + mean+std pool, 4,491-dim pre-PCA), used in the SI Section S1.5 robustness ablation; 75.2% of variance retained. crystal_communities_artifacts.tar.zst — all per-claim run outputs, hierarchically organized under work2_snapshot/. Includes the canonical labeled Louvain partition, time-evolution analysis, formula stepping-stone (the source for the 82.9% number cited in the manuscript), composition-matched 5-source control, held-out 1990/2000/2010 cutoffs, A-Lab outcome cross-validation, synthesis pivot, functional-class labeling, the chemistry/structure ablation study, the k/Louvain-resolution sensitivity sweep, the 90/95/99-percentile threshold sensitivity, the theoretical-flag re-run, and the methods-text-faithful ablation. Decompress with tar --zstd -xf crystal_communities_artifacts.tar.zst. README.md — full layout, exclusions, and use instructions. CHECKSUMS.md5 — md5 sums recorded by Zenodo at upload time, for integrity verification. Excluded by design The licensed ICSD CIFs themselves (ICSD_CIFs.zip, FIZ Karlsruhe). Not redistributable. The featurization in this deposit replaces the need for raw CIFs for any downstream analysis. To re-derive the embedding from CIFs, run the pipeline in the GitHub repository against your own licensed ICSD copy. Abandoned exploratory runs (e.g. geometry_only_ablation/) not cited in the paper. The 5.7 GB raw features.npy from the methods-text-faithful ablation — re-derivable by re-running the ablation pipeline; only the 41 MB PCA-reduced version is preserved. Conda environments (~6 GB total) — recreate from environment.yml in the GitHub repository. Reproducibility The student/reviewer code guide in the GitHub repository (notes/code_guide.md) gives one-liner reproduction recipes pointing at the artifacts in this deposit, including worked examples for the 82.9% TRI-restricted stepping-stone rate, the 76.8% PCA(32) explained variance, the partition-agreement ablation (ARI 0.49, NMI 0.89 vs the methods-text-faithful variant), and the per-source held-out frontier rates. Every load-bearing claim of the manuscript can be verified from this deposit alone, without requiring ICSD access. Citation Please cite the manuscript (DOI to be added on publication) alongside this deposit’s DOI. B. Chu's contribution to early structural-similarity prototyping is acknowledged in the manuscript.