The molecular landscape of CPOSS. The horizontal and vertical coordinates are dimensionless sketch-map coordinates; their orientation is arbitrary and neither axis corresponds to an individual molecular property. Each circle represents one CSP dataset (307 of 310 datasets; 517,431 archived structure files represented). A representative crystal was selected from the lowest-numbered documented energy-model archive with a reconstructable molecule, using the lowest finite energy printed in its TITL field after excluding explicit experimental-reference labels; up to five candidates per model were validated. Complete molecules were reconstructed using SHELX symmetry and periodic covalent connectivity. The largest non-water component was retained, with coformers, counterions and stoichiometry recorded separately. The map compares L2-normalized, outer-averaged SOAP power spectra (all atoms and elements; cutoff 4.5 A, Gaussian width 0.4 A, n_max=6, l_max=4; 11,055 features). Molecular distances are sqrt(2-2K), with K the squared normalized descriptor dot product. scikit-matter sketch-map minimizes sigmoid-transformed pair-distance stress with sigma=0.4968, high- and low-dimensional exponents (1,4), and equal dataset weights. Five initializations were tested and the lowest-stress solution was refined. Marker area is 1.8 sqrt(N) square points, where N is the number of model-specific RES files across the dataset's documented archives; the square-root compression keeps small datasets visible. Counts do not weight the embedding and have not been geometrically deduplicated. Colour encodes the extracted component's molar mass on a logarithmic truncated-Plasma scale. Ball-and-stick callouts use matte VMD-inspired materials and native Tachyon ambient-occlusion ray tracing from the same extracted coordinates used in the map. Eight recognizable, experimentally characterized compounds were curated: urea, succinic acid, aspirin, paracetamol, caffeine, carbamazepine, naproxen, chlorothalonil. Their CPOSS experimental crystal references establish compound identities, not experimental observation of the selected predicted packing or conformer. Standard element colours are used. Leader lines terminate at unjittered dataset positions. Repeated identities and coincident points remain separate dataset records and can overlap. Dark outlines flag 11 source-formula conflicts, retained without editing source formulas. 3 inconsistent element tables prevented molecular assignment and are excluded. Bond orders could not be resolved for 3 formula-consistent components; their SDFs are connectivity-only, and SOAP uses their nuclear coordinates. SOAP does not resolve enantiomers, crystal packing or omitted coformers; proximity reflects composition and local molecular geometry rather than a unique chemical classification.
