NOCTIS ARRAY archive
The array is not.
A BLACK HOLE LEARNING TO GLOW.
Every difference image our cameras produce contains two kinds of change: something that moved, and something that didn't. An active galactic nucleus is the second kind — a point of light sitting exactly where a galaxy's core has always been, brightening and dimming for reasons that have nothing to do with proper motion. This page is about how to tell that flicker from a moving asteroid, a pulsating star, or plain noise — and, honestly, about how well the real pipeline behind this site actually does at it.
NOCTIS ARRAY does not exist. Active galactic nuclei do — and this page explains them the way a real time-domain difference-imaging pipeline actually has to think about them.
A supermassive black hole, caught in the act of feeding.
Every large galaxy appears to carry one at its centre. Most sit quiet, starved of fuel. An active galactic nucleus is what happens when one is not.
A black hole a few million to a few billion times the Sun's mass, pulling gas into a spiralling accretion disk hot enough to outshine every star in its galaxy combined.
The engineAccretion, not fusion.
Gas falling toward the black hole doesn't fall straight in — it spirals into a flattened accretion disk, friction and turbulence heating it to hundreds of thousands of degrees. Converting infalling mass into radiation this way is dramatically more efficient than the fusion powering an ordinary star: up to tens of percent of rest-mass energy, against roughly one percent for a star.
That efficiency is how an object smaller than a solar system can, for a while, outshine a hundred billion stars.
The unified modelFour names, one object.
Seyfert galaxy, quasar, blazar, radio galaxy — these read like four kinds of object. Under the unified model they are mostly one: the same accreting black hole seen at different luminosities, different viewing angles through a surrounding dusty torus, and with or without a relativistic jet pointed our way.
A Seyfert is the same engine at modest luminosity, its host galaxy still visible around it. A quasar is the same engine turned up, often outshining its host entirely. A blazar is a jet aimed almost straight at Earth, boosted by relativity into wild, rapid variability. A radio galaxy is the same jet, viewed from the side. Classification often says more about geometry than about physics — though not entirely: luminosity and accretion rate genuinely differ between these objects too, and a handful of "changing-look" AGN switch type outright, which pure orientation can't explain.
Why it flickersNo two nights the same.
The accretion disk is turbulent, not smooth — clumps of gas fall in unevenly, and the whole system has no preferred rhythm. The result is stochastic variability: real, repeated brightness changes with no fixed period, closer to a random walk than a heartbeat.
That is the opposite of a pulsating star, which brightens and dims on a clock precise enough to set a distance scale by. The absence of a period is itself information.
Five ways a nucleus gives itself away.
None of these alone is proof. Together — position, light-curve shape, colour, spectrum and wavelength — they build a candidate a human, or a classifier, can trust.
-
01
Position — nucleus-coincident, and it doesn't move.
An AGN candidate sits exactly on a galaxy's core, exposure after exposure. That fixed position is what separates it immediately from an asteroid or a satellite streak — the entire class of object this pipeline's motion branch exists to catch instead.
-
02
Light-curve shape — stochastic, not periodic.
Repeated difference-imaging visits build a brightness history over time. A pulsating or eclipsing star repeats on a clock; an AGN wanders — brightening and dimming with no fixed period. That aperiodic wander is the signature a time-domain survey is built to catch across many revisits to the same patch of sky.
-
03
Colour — bluer than the galaxy around it.
An accretion disk's continuum tends to run bluer than a galaxy's ordinary starlight, so an AGN candidate often shows a blue or ultraviolet excess relative to what its host galaxy's colour alone would predict.
-
04
Spectroscopy — broad emission lines, when a spectrum is available.
Gas orbiting close to the black hole moves fast enough that its emission lines are Doppler-broadened into wide humps. Seeing those broad lines — an unobscured, "type 1" view — is close to definitive confirmation. A fully obscured "type 2" nucleus shows only narrow lines and needs another method entirely.
-
05
Multiwavelength counterpart — X-rays and radio, where the sky has been watched.
A hot accretion-disk corona is a strong hard-X-ray emitter, and a jetted AGN can be loud in radio. A point source at a galaxy's nucleus that also carries an X-ray or radio counterpart adds independent evidence a purely optical survey can't provide on its own.
Signal 1 and signal 2, drawn instead of described — and why signal 2 alone is a weak one:
What our real pipeline actually does with an AGN candidate.
Every AGN mention above stops being NOCTIS fiction the moment it touches this box. This repository's real ZTF pipeline runs a pretrained ALeRCE stamp classifier — a single-image, TF1-era CNN checkpoint, run locally, no training performed here — on detections that already passed a separate real/bogus gate. It sorts survivors into four classes: AGN, SN (supernova), VS (variable star), and asteroid.
On our own detections, one real ZTF source, src_p1_024, was typed AGN at P = 0.549. Run across a full field of 114 real 2020 ZTF detections, the classifier called 12 of them AGN (against 45 bogus, 36 VS, 21 uncertain) — a smaller but genuine share, roughly what a typical stationary star-and-galaxy field should contain.
On a fair labelled evaluation — native ALeRCE stamps, real metadata, the correct four classes — that same local classifier reached about 0.70 overall accuracy, but was weakest specifically at telling AGN apart from variable stars (≈0.55 accuracy). That is a genuine, documented astrophysical ambiguity, not a bug: both are point sources sitting at a fixed position, and a single image at a single epoch often can't tell "a flickering AGN nucleus" from "a variable star" apart.
A separate pretrained light-curve random-forest classifier files AGN-like variability under its top-level "Stochastic" class, distinct from "Periodic" for pulsating or eclipsing stars. In this project's own light-curve evaluation, AGN/QSO sat among the weaker classes — the two strongest were both periodic star types, eclipsing binaries and RR Lyrae, not AGN. The pipeline has classified known ZTF objects with real, documented uncertainty. It has not discovered anything, and does not claim to.
A flickering nucleus and a flickering star can look identical — once.
The signature list above is a five-way check. In practice, a real pipeline rarely gets all five on any one candidate, and the honest gaps are worth naming.
One photograph of a point source that has changed brightness tells you almost nothing about why.
Single image, single epochTwo point sources, same evidence.
A galaxy-nucleus AGN and a variable star both show up in a difference stamp as a compact source that brightened, sitting at a fixed position, with no motion between exposures. Shape and position alone don't separate them — which is exactly the AGN-vs-VS confusion this project's own classifier evaluation documents at ≈0.55 accuracy.
Colour and spectroscopy help when they're available. A single-band difference image, on its own, usually isn't enough.
What a light curve is supposed to addA period settles it — when there's enough of one.
Repeated revisits should separate the two: fit enough of a light curve and a periodic pulsator reveals its clock, while an AGN's stochastic wander stays stubbornly aperiodic. That is the theory behind the light-curve classifier's "Stochastic" branch.
In practice, this project's own light-curve evaluation found the periodic classes (eclipsing binaries, RR Lyrae) easier to nail than the stochastic ones — AGN/QSO included — largely because sparse, alert-stream sampling gives a clean period more to grab onto than it gives a noisy random walk. The harder case stayed hard.
No discovery is claimed, here or anywhere on this site. Everything in the box above describes our pipeline classifying already-known ZTF objects, with real, documented uncertainty attached to the numbers. NOCTIS ARRAY, its site, its timeline and its own invented statistics are fiction, exactly as the About page discloses — the astrophysics on this page, and the results in the callout above, are not.
A feeding black hole and a pulsing star can produce the same one photograph.
Both are a compact source at a fixed position that got brighter or fainter. Here's what actually breaks the tie — first the cues a person would reach for, then the numbers our own classifiers actually compute to reach for the same thing.
-
01
What's underneath it — a smudge of galaxy, or a bare point?
An AGN sits at the centre of a whole galaxy — if that host is resolved in the image, a soft, diffuse glow surrounds the bright nucleus. A variable star is a naked point source with nothing extended around it, because it's just one star. This is the single most reliable tell when it's available, and the single most useless one when it isn't: a distant or faint AGN's host blurs into an unresolved point, indistinguishable from a star, at exactly the resolution most surveys actually deliver.
-
02
The shape of its light curve — a wander, or a clock.
This is the one that actually decides most cases. An AGN's accretion disk has no preferred rhythm — its brightness is a random walk, not a cycle, so no period ever makes the points line up. A pulsating or eclipsing star repeats: fold enough visits on the right period and every point falls back onto the same curve. The catch is coverage: this tell only works once there are enough revisits to either confidently rule out a period or find one, and a sparse light curve looks the same — uninformative — either way.
-
03
Colour and amplitude — a hint, not a verdict.
An accretion disk's continuum tends to run bluer than ordinary starlight, and an AGN's swings can be large and abrupt rather than the smooth, bounded rise-and-fall of a pulsator. Both are soft cues on their own — plenty of hot, blue stars exist, and plenty of AGN sit quiet for long stretches — which is exactly why no single signal above is treated as proof by itself, on this page or in the pipeline below.
How the ML actually looks for the same three things. Neither classifier "sees" a galaxy or a star the way a person does — each one turns the tells above into plain numbers.
The stamp classifierreads "extended vs. point" as a number.
Each detection is scored from a small image cutout plus a metadata vector of about two dozen real measurements — among them a Pan-STARRS point-source score for that position and its nearest neighbours. That score is the numeric stand-in for "does this sit on something extended, or is it a bare point" — the same host-structure cue from tell 01 above, just handed to the network as a number instead of a picture.
The light-curve classifiersplits on stochastic-vs-periodic first.
Roughly 180 numeric features get computed from each object's brightness history — period search results, amplitude, how lopsided the scatter is — and fed to a pretrained hierarchical random forest whose very first split is Stochastic vs. Periodic vs. Transient. That top-level branch is tell 02 above, made numerically: a shapeless wander pulls an object one way, a real period pulls it the other.
Both features exist and both are real. They just run out of information at the same place a person's eye would — a faint, sparsely-sampled point source with an unresolved host and too few visits to call a period either way. That's the ≈0.55 accuracy documented above, explained mechanically rather than just stated.
Go look at how it actually classified them.
The array above is invented. The browsers below are not — they open a working ZTF difference-imaging pipeline rendering real observations, real detections, and real classifier scores, AGN included.
Back to the overview