Fig. 5: Visualization and exploration of the latent space.
From: Property-guided generation of complex polymer topologies using variational autoencoders
a Two-dimensional visualization of the TopoGNN latent space using the Uniform Manifold Approximation and Projection (UMAP) technique. A subset of the data is displayed for clarity, with each marker representing a polymer graph based on its latent vector. Different colors denote distinct topologies. b Organization of (left) \(\langle {R}_{{{{\rm{g}}}}}^{2}\rangle\) and (right) topology in the UMAP-coordinate space. The dots signify the latent vectors of polymer graphs. The two arrows mark regions in the latent space targeted for exploration (i.e., new polymer topology generation). c As exploration progresses with an increase in Z2 in the latent space (represented by a solid line), there is a near-monotonic rise in \(\langle {R}_{{{{\rm{g}}}}}^{2}\rangle\) for the generated polymers. d Progression with an increase in Z1 (indicated by a dashed line) showcases shifts in polymer topology, moving through clusters characteristic of star, comb, and branch topologies.
