Spectral Twins (DAEP)

Explore the latent space of a diffusion autoencoder (DAEP) trained on supernova spectra. Select a spectrum to identify its spectroscopic twins through latent space nearest-neighbor searches using cosine similarity, compare their spectra, and examine the model's reconstructions.
Model and data: Henna Abunemeh (UIUC).

Loading the spectral twins data (about 69.4 MB before compression)…
Spectroscopic twins are identified using cosine similarity/distance between full, 1024 dimensional DAEP latent representations. A 2D UMAP projection is constructed from the precomputed distance matrix, UMAP is only used to visualize the latent space structure.