Decoding the Unseen: Unsupervised Anomaly Detection in Metal–Organic Frameworks for Discovery Beyond the Norm

The discovery of chemically novel or structurally anomalous metal–organic frameworks (MOFs) is essential for expanding reticular design space and improving dataset reliability. We present CHEM-AD (Chemically Unusual Metal–organic Frameworks via Autoencoder-based Detection), a label-free, CPU-efficient pipeline that detects anomalous MOFs using 32 standardized geometric, chemical, and topological descriptors. A compact symmetric autoencoder (32–16–8–16–32; 1,352 parameters) learns a latent manifold of typical structures and assigns reconstruction-error scores to each entry. Applied to 26,025 MOFxDB records, the score distribution is right-skewed; a knee on the sorted curve at τ = 0.4666 flags 729 candidates (~2.8%). These candidates are not random artifacts: they occupy peripheral islands and filament endpoints in t-SNE/UMAP and also exhibit large Mahalanobis distance. The top-10 span PLD = 2.56–29.48 Å, LCD = 4.89–63.59 Å, density = 0.0569–4.272 g cm⁻³, and surface area = 141–8,318 m² g⁻¹; all fail a topological sanity check while roughly half pass chemical and geometric checks—implicating connectivity (node/edge counts, component statistics, degree dispersion, transitivity) rather than any single textural extreme as the dominant driver. We organize outcomes into a practical triage: (A) topologically uncommon yet chemically/geometrically plausible; (B) the same signature with minor chemistry flags resolvable by bond–valence/charge-balance checks; and (C) records best routed to CIF re-processing. CHEM-AD requires no graph extraction or topology labels and runs reproducibly on commodity CPUs, enabling rapid discovery, dataset curation, and robust preprocessing. Datasets (26,025×32 descriptor matrix; per-MOF anomaly scores; per-feature errors) are released on Zenodo; code is available on GitHub.