256,000 coordinate realizations span 24 design groups. The original 64,000 are preserved exactly; 192,000 additions broaden the original families and introduce eight hybrid motif groups. Colors identify design groups. The descriptors and embedding use geometry only: labels and mechanical properties do not enter the fit.
Axes have no physical units. UMAP approximates local geometric similarity; cluster shapes and global distances are not material laws. Nearby thumbnails use distance in this displayed map.
These are newly generated, unrelaxed structures, not 256,000 simulations from the original paper. Atoms lie in a plane at z = 10 Å, in a cell with Lz = 20 Å and periodicity x/y only. The vacuum cell height is not graphene thickness. Edges are unpassivated; mechanical stability requires separate calculations. Graph connectivity is available as a geometric screening measure.
The population switch compares the original and added designs within the same newly fitted map. Candidate selection favors descriptor novelty and cross-group interpolation queries; empty regions of UMAP are not automatically feasible missing designs. Periodic path rank is the number of independent winding cycles in the initial neighbor graph, not a mechanical validation. Distinct coordinate realizations do not necessarily imply distinct topology. No energies, forces or mechanical labels are supplied.
Dataset documentation, code and validation ↗
Apache 2.0 · LAMM, MIT