Self-supervised transformer (zsig) on the Chen+2020 catalog (~25 K periodic variables, 11 types). Click a point to see its raw or phase-folded light curve and Gaia HR-diagram position.
A representative source per variability class. Click to jump.
Click a point in the UMAP scatter to inspect it. The right column
panels each have a × button to collapse and a ⛶ button (LC only)
to maximize. URL hash #<idx> deep-links to a
specific source.
Logistic-regression probe on z_sig (5-fold CV). Click any cell to open the sources that fell in (true row × predicted column).
Pre-trained self-supervised on raw multi-band ZTF light curves with InfoNCE
contrastive loss on z_sig and an auxiliary regression on
z_qual to encourage disentanglement. Each contrastive pair
uses two disjoint random time-windows of the same source so the model
can't shortcut the objective by matching cadence. The 25K Chen sources
used to compute the probe accuracy below are held out from the training
pool. Use the Layer dropdown above to UMAP each encoder layer's
pooled output and see when class structure emerges.
ZTF visits the same patch of sky at irregular times — every few nights, gaps in bad weather, blocked by the Sun for months. Watching the data accumulate in real time (left) it looks like noise. Folding each observation onto the same period (right) reveals the periodic light curve hiding underneath.
Example: an RR Lyrae pulsator — surface oscillates with P = 0.539 d, 196 ZTF observations spanning ~3 yr. Same physical brightness sequence, two ways of looking at it.