AI in Astronomy: Classification at Scales Humans Cannot Reach

Astronomy adopted machine classification early and without much argument, for an unglamorous reason: the data arrives faster than anyone can look at it, and it always has. The interesting part is not that the classifiers work. It is what a working classifier does to the statistics you compute afterwards.

The firehose

A modern time-domain survey photographs the sky repeatedly and looks for what changed. The standard method is difference imaging: align a new exposure with a deep reference image, subtract, and see what is left. Anything genuinely new or variable shows up in the residual — and so does everything that went wrong. Imperfect alignment leaves dipole artefacts around every bright star. Cosmic rays hit the detector. Satellites draw streaks. Bad columns, saturation spikes and convolution artefacts all produce residuals that look, to a threshold, exactly like a source.

The result is a stream of candidate detections in which the artefacts vastly outnumber the astrophysics. Surveys now generate these alerts at rates in the millions per night, and the next generation of wide-field instruments pushes it further. There was never a version of this where humans looked at each one, so the first classifier in the chain — real against bogus, run on small image cutouts — is not a research convenience. It is the only reason the stream is usable at all, and it has been operational infrastructure for years. On the ladder in the cluster overview this is the top rung: practice changed, permanently.

What gets classified

• Transient typing from light curves. Deciding what kind of explosion or variable a new source probably is, from its brightness over time and colour, so that scarce spectroscopic follow-up time goes to the interesting ones. The value here is allocation of a genuinely limited resource: telescope hours.
• Galaxy morphology. Shape classification at catalogue scale. This is the field where large crowdsourced labelling campaigns produced training sets first, and supervised models then scaled the labelling to surveys no volunteer effort could cover.
• Rare object search. Strong gravitational lenses are visually distinctive and extremely rare, which is the worst possible combination for manual searching and a natural fit for a classifier that reduces millions of cutouts to a reviewable candidate list.
• Transit vetting. Exoplanet searches produce far more candidate signals than planets, most of them instrumental or eclipsing binaries. Applying convolutional classifiers to archival space-telescope light curves surfaced candidates earlier vetting had set aside, which then went through the field’s standard statistical validation before anyone called them planets.
• Instrument artefact classification. Gravitational wave detectors produce characteristic noise transients, and sorting them by morphology is how the instrument gets debugged. Note the division of labour: detection of the signals themselves rests on matched filtering against modelled waveforms, and the learned component is used around it rather than instead of it.

The selection function is the result

Here is the methodological point that separates a demonstration from a publishable result, and it generalises far outside astronomy.

Suppose you want to know how common some class of object is, or how its rate varies with distance. You count what is in your catalogue. But your catalogue is what survived the pipeline, and the pipeline includes a classifier that rejects things — with an efficiency that is not uniform. It probably recovers bright, slow, well-sampled examples more often than faint, fast, badly-sampled ones. If you do not know that function, your measured rate is the true rate multiplied by an unknown quantity that varies with exactly the properties you are trying to study.

So the classifier’s completeness and contamination have to be measured as functions of the object’s properties, and the standard way to do it is injection and recovery: insert synthetic sources with known properties into the real data, at the pixel level, upstream of everything, and run the whole pipeline. The fraction recovered, as a function of brightness and every other relevant parameter, is the selection function. It is a substantial piece of work and it is not optional — without it a population number is a statement about a neural network, not about the universe.

The general form: any time a learned filter sits between the world and your dataset, characterising the filter is part of the measurement.

When the likelihood is intractable

The second major use is different and grows out of a structural feature of the field. Astronomy has excellent forward models: given physical parameters, simulate what an instrument would record. What it often lacks is a tractable likelihood — a formula for the probability of the observed data given the parameters — because the forward model involves simulation steps with no closed form.

Simulation-based inference attacks this directly. Sample parameters, run the forward model, and train a network on the resulting parameter-and-data pairs to approximate the posterior. Once trained, inference on new data is fast, which matters for anything that has to respond to an alert while the event is still observable. Structurally this is a surrogate for the inference step rather than for the physics, and it carries the same caveat in a sharper form: the posterior it learned is the posterior implied by the simulator. If the simulator is missing a systematic, the network will confidently report a tight, wrong credible interval. This is why coverage tests — checking that stated intervals actually contain the truth at the stated rate on held-out simulations — are treated as mandatory rather than as a nicety.

Where classifiers are weakest

• Exactly at the discovery frontier. A supervised classifier is trained on known classes. Something genuinely new resembles none of them, and the most likely outcome is that it is assigned to the nearest familiar category or rejected as an artefact. This is why unsupervised anomaly detection is run alongside classification rather than replaced by it.
• Labels inherited from older pipelines. Training sets are frequently built from previous catalogues, which had their own selection effects. The new classifier then learns to reproduce them, including the mistakes, and the agreement looks like validation.
• Domain shift between instruments. A model trained on one survey’s images encodes that telescope’s point spread function, depth and cadence. Transfer to another instrument is a distribution shift, and performance measured on the first says little about the second.
• Rarity ruins accuracy as a metric. When the positive class is one in a million, a classifier can be accurate to five decimal places and useless. The quantities that matter are recall at a fixed, affordable follow-up budget, and the purity of the list a human actually reviews.