Reading a galaxy's redshift from the filters it vanishes from.
Neutral hydrogen across the early Universe absorbs a galaxy's light blueward of the Lyman break (1216 Å). As a galaxy sits further away its light is stretched, so that break sweeps redward through the filters, and the galaxy drops out of the bluer bands one by one. The pattern of which filters it appears in is its redshift. A tiny neural net, exported to ONNX and running live in your browser, reads that pattern off nine JWST/HST bands. This is how JWST pushed the frontier to z ≈ 14 and turned up too many, too bright early galaxies. loading model…
Select a galaxy below.
Every dot is a real CEERS galaxy, placed by two colours and tinted by redshift. As the break crosses F115W, high-z galaxies redden sharply in F115W−F150W and rush to the right, into the dropout corner. The selected galaxy is ringed.
What you're seeing. Real JWST/NIRCam cutouts (and HST/ACS where it helps), pulled from the DAWN JWST Archive cutout service for galaxies in the CEERS field. Fluxes and reference redshifts come from the public grizli / EAZY photometric catalogue (van der Wel et al. 2025). The network was trained on ~71,000 template redshifts and is scored here against ~1,800 real spectroscopic redshifts it never saw.
Step a galaxy up in redshift: the dropped-out band marches from F115W toward the red. The marker on the wavelength bar is the Lyman break for that redshift.
why it drops outIntergalactic neutral hydrogen absorbs almost everything blueward of 1216 Å rest-frame. Redshifted into the observed bands, that wall of absorption blacks out every filter sitting below it.
So the bluest filter a galaxy still shines in pins down (1+z).
A few candidates give a two-humped PDF: a dusty galaxy whose 4000 Å break mimics a Lyman dropout. The net shows the doubt instead of hiding it.
the great degeneracyA red z ~ 2 galaxy and a z ~ 11 dropout can show nearly the same colours. This ambiguity is exactly why JWST's "too many bright early galaxies" needs spectroscopy to confirm, and why some candidates evaporate.
Accuracy is best at z 4–6, where the break sits squarely inside the filters; at low z it lives in the unobserved UV and the net is hazier.
distilled, not magicThe net is distilling template fitting into something that runs instantly client-side. It reaches σNMAD ≈ 0.04 against spectroscopy, within a hair of the templates' own 0.027, and at a millionth of the cost.