CHEM-AD: MOFxDB descriptors and anomaly scores (v2.0)

Description We release the data underlying CHEM-AD, a CPU-efficient, label-free anomaly detector for metal–organic frameworks (MOFs). The model operates on 81 engineered descriptors per MOF (32 geometric/chemical/topological scalars plus a 49-dimensional metal-composition encoding). Scope.26,025 MOFxDB entries. Anomaly scores come from a symmetric autoencoder trained with MSE on standardized descriptors (input dimension 81, 16-dimensional latent space, ∼1.8×10⁵ trainable parameters). A knee threshold of τ ≈ 1.6882 on the sorted scores flags 488 anomalies (≈1.87% of the dataset).