Supernovae
Recurrence never
A STAR THAT VARIES ONCE.
Every other kind of change this array records is something a star will do again. A supernova is the exception: a star's actual death, bright enough for a few weeks to outshine the entire galaxy it sits in, and then gone for good. That makes it the cleanest thing a difference image can possibly show — a point of light where the reference frame holds nothing at all — and simultaneously the hardest to catch, because you have to already be looking. The instrument below is invented. The physics is not.
NOCTIS ARRAY does not exist. Supernovae do — and this page explains them the way a real time-domain difference-imaging pipeline actually has to think about them.
Two completely different accidents, one similar-looking result.
"Supernova" is not one event. It is two unrelated physical catastrophes that happen to release comparable energy over comparable timescales, which is why they were classified together for decades before anyone knew they were different things at all.
One is a bomb going off. The other is a building falling in on itself.
Thermonuclear · Type IaA white dwarf detonates.
A white dwarf is the exposed carbon–oxygen core of a dead star, held up not by heat but by electron degeneracy pressure — a quantum-mechanical floor rather than a thermal one. That floor has a hard limit near 1.4 solar masses. Push a white dwarf past it, by stealing gas from a close binary companion or by merging with a second white dwarf, and carbon fusion ignites across the whole star essentially at once. There is no core left behind; the star is entirely unbound.
Because the trigger is a fixed mass threshold, Type Ia explosions are broadly alike, and their peak brightness correlates tightly with how fast they fade. That relation is what turned them into standardisable candles — the distance measurement that revealed the accelerating expansion of the universe. Spectroscopically they are recognised by what is missing: no hydrogen, because the progenitor shed it long ago, with strong silicon absorption instead.
Core collapse · Types II, Ib, IcA massive star runs out of fuel.
A star above roughly eight solar masses fuses its way up the periodic table, each stage faster than the last, until it builds an iron core. Iron fusion consumes energy rather than releasing it, so the core's support vanishes in under a second. It collapses to a neutron star — or a black hole — and the infalling outer layers rebound off that new surface, with an enormous neutrino flux helping drive the shock outward.
The subtypes are a stripping sequence, not different mechanisms. Type II still has its hydrogen envelope, so hydrogen lines dominate its spectrum. Type Ib lost the hydrogen. Type Ic lost the helium too. Type II light curves often show a months-long plateau as the expanding hydrogen shell recombines and releases stored energy at a nearly constant rate — a shape a survey can recognise from photometry alone.
What powers the glowRadioactivity, not the blast.
The explosion itself is over in seconds. What a survey actually watches for weeks afterwards is radioactive decay: the blast forges roughly half a solar mass of nickel-56, which decays to cobalt-56 and then to stable iron-56, dumping gamma rays into the expanding debris cloud that thermalise and escape as visible light.
That is why supernova light curves have the shape they do — a rise of a couple of weeks as the ejecta expand and more of that trapped energy escapes, then a decline whose slope follows the cobalt-56 half-life. The clock is nuclear physics, which is precisely why the light curve is diagnostic rather than arbitrary.
The one case difference imaging was practically designed for.
A variable star is a source that was already there getting brighter. A supernova is a source that was not there at all. Subtract a deep reference built before the explosion and what remains is a clean, isolated point of new light with nothing underneath it to confuse the measurement.
No residual to cancel, no host to subtract off. Just light that did not exist last month.
Look where the galaxies areTargeting beats luck.
Supernovae happen in galaxies, because that is where the stars are. A blind wide-field survey finds them anyway by covering enough sky often enough, but the detection is always the same shape: a new point source within, or just outside, the visible extent of a host galaxy — usually offset from the nucleus rather than sitting on it, because the exploding star was somewhere in the disc, not at the centre.
Catch the rise, not the peakCadence is the whole game.
The scientifically valuable part of a supernova is the first few days, when the shock breaks out and the ejecta are still compact — that early behaviour constrains the progenitor star's size and the explosion's geometry. Peak brightness comes two or three weeks later and tells you much less. A survey revisiting every couple of nights can catch the rise; one revisiting monthly finds only the decline.
Then hand it offPhotometry classifies, spectra confirm.
Photometry alone — brightness versus time, in a couple of filters — is genuinely good at separating a transient from a periodic variable, and reasonably good at guessing the subtype from light-curve shape and colour. Confirming which supernova it is, though, means a spectrum, from a telescope that has to be pointed at it deliberately. Surveys are the alarm; the classification is somebody else's night.
What that subtraction actually produces for a transient, drawn instead of described:
Supernovae are the rarest label our real pipeline has ever assigned.
Across 666,811 detections in 503 processed ZTF exposures, the classification stage has typed exactly 15 detections — five distinct sources — as SN. For comparison, the same runs produced 15,541 variable-star detections. Supernovae are the smallest astrophysical class in the whole database by two orders of magnitude.
Three separate reasons stack up, and only one of them is about the software. First, supernovae genuinely are rare — roughly one per galaxy per century, so a patch of sky has to be watched a long time or contain a great many galaxies. Second, our fields are ordinary equatorial star fields chosen for archival depth, not galaxy-cluster fields chosen for transient yield: they are dominated by foreground stars, which is exactly the population the VS number reflects. Third, and most importantly, our object type comes from cross-matching each detection against ALeRCE's catalogue of alerted ZTF objects — so the pipeline can only put an SN label on something the alert stream already knew about.
Where supernovae do stand out is model performance. On a fair labelled evaluation of the local stamp classifier — native ZTF cutouts, real alert metadata, objects genuinely in its four classes — SN scored 0.80 — well ahead of variable stars at 0.60 and AGN at 0.55, and second only to asteroids at 0.89. That is not a coincidence: a new point source on a galaxy with nothing beneath it in the reference is a far more distinctive picture than one more unremarkable point that got brighter. The separate light-curve classifier puts supernovae under its top-level Transient branch, the one class of object it does not try to find a period for.
None of these are discoveries. Every one of the five sources was an already-catalogued ZTF alert; the pipeline recognised it, it did not find it first. What is real here is the method, the counts and the 0.80 — the observatory drawing them onto this page is not.
A brand-new point of light has several unglamorous explanations.
The clean picture two sections up is the ideal case. On a real night, three other things produce a positive residual at a position with no reference counterpart, and none of them is a supernova.
One epoch can tell you something new appeared. Only a second epoch can tell you what.
An asteroid drifted acrossNew light that moves.
A main-belt asteroid passing over a galaxy produces exactly the same signature in a single frame: a point source with no reference counterpart. The difference only shows up on the next visit, when the supernova is still at the same coordinates and the asteroid has moved on. This is why nothing on this site treats one detection as an object, and why the asteroid page next door exists.
A cosmic ray hit the detectorNew light that has no shape.
A charged particle striking the CCD directly deposits its energy in one or a few pixels with no point-spread function at all — unnaturally sharp compared to any real star, and present in a single exposure only. That sharpness is precisely what the real/bogus classifiers are trained to notice, and it is why our pipeline runs one before it ever tries to assign a type.
The nucleus flared insteadNew light that is not new.
An active galactic nucleus brightening looks like a transient on a galaxy too — except it sits on the nucleus rather than offset from it, and it will do it again. Position relative to the host is the fast tell; a light curve that keeps wandering instead of decaying on a cobalt-56 slope is the decisive one.
And the failure mode nobody sees. If the deep reference used for the subtraction was built from epochs that already contained the supernova, the transient partly cancels itself out and the residual is weak or absent. This project hit exactly that wall: an early four-epoch reference spanning about a month self-subtracted its own variable sources, which is why the reference was rebuilt from forty frames spread across several years. Nothing errors when this happens — the source simply is not there. It is the strongest argument on this page for a reference that predates what you are looking for.
Three classes, one photograph, three different ways to break the tie.
A supernova, an active galactic nucleus and a variable star can all present as a compact source that changed brightness. Here is what actually separates them — and what our own classifiers compute to reach for the same distinctions.
-
01
Is there a counterpart in the reference at all?
This is the supernova's single strongest tell and the one the other two cannot fake. A variable star was already there and merely changed; an AGN was already there and merely changed. A supernova has no reference counterpart whatsoever, so the difference image shows new flux over blank sky (or over a host galaxy that itself cancels cleanly). The stamp classifier gets this for free: it is fed the science, reference and difference cutouts as three channels of one image, so "bright here, empty there" is directly visible to the network rather than something it has to infer.
-
02
Where is it, relative to the galaxy underneath?
An AGN is the nucleus — by definition it sits at the host's centre. A supernova sits wherever the progenitor star happened to be, which is almost always offset into the disc. A variable star usually has no host at all. The metadata vector our stamp classifier consumes alongside the image carries roughly two dozen real measurements, including a Pan-STARRS point-source score and the distances to the nearest catalogued neighbours — the numeric stand-in for "what is this thing sitting on, and where on it".
-
03
Does the light curve come back down and stay down?
This is the decisive one once enough visits exist. A supernova rises for a couple of weeks and then declines on a slope set by radioactive decay, once, forever. A variable star repeats on a period. An AGN wanders with no period at all. That is precisely the three-way split at the top of the light-curve classifier's hierarchy — Transient, Periodic, Stochastic — computed from roughly 180 numeric features of the brightness history rather than from any single image.
Why the classifier needs at least six visits. The light-curve model in this codebase refuses to run on an object with fewer than six detections in the g or r band, and that refusal is the honest version of tell 03.
Below six pointsevery class looks the same.
Three scattered measurements can be fitted by a decline, a period or a random walk with equal ease. Rather than guess, the model declines to answer — which is why so few of our own field's faint detections ever reach it, and why the light-curve evaluation had to be run against alert-rich catalogued objects instead.
Above six pointsthe transient branch separates cleanly.
A monotonic decline is a genuinely distinctive shape once you have enough of it. That is why the same evaluation that scored 0.38 on fine classes overall still separated the top-level Periodic / Stochastic / Transient split at 0.66 — the coarse question is the answerable one.
Go see what it actually classified.
The array, the cadence and the field notes above are invented. The browsers below are not — they open the same working ZTF difference-imaging pipeline whose five SN-typed sources are described in the panel above.
Back to the overview