Clustering in astronomy and astrophysics

How it is used

Astronomy produces catalogues of objects described by many measurements at once: brightness in several colours of light, position, motion, chemical makeup. Clustering is a family of methods that sort such objects into groups by how close their measurements are to one another, without being told in advance what the groups should be. Related techniques squeeze many measurements down to a few summary numbers first, so that the grouping happens in a simpler space. The same idea works on positions on the sky: points that fall close together can be gathered into one detection, separating real pile-ups of points from scattered stragglers.

In practice the method is used to divide a sample into populations that the physics suggests should exist, such as two chemically distinct parts of our galaxy, and to merge many independent marks on an image into a single object. It is also used earlier in a study, as preparation: deciding how to rescale or transform measurements so that distances between objects mean something sensible.

Why researchers use it

Astronomers reach for clustering because the sky is observed, not arranged. Nobody hands over a list of which star belongs to which galactic population, or which smudge on an image is one object rather than three. Labels are scarce and expensive, usually requiring follow-up observation, while the raw measurements arrive in bulk. Clustering stands in for that missing labelling: it proposes the groups from the data itself, which can then be checked against what is known about the physics.

It also stands in for human sorting at a scale humans cannot reach. Archives hold more images, and surveys more objects, than anyone can go through by hand. Where people do contribute, as with volunteers marking features on pictures, clustering turns a scatter of individual judgements into one agreed answer per object, and quietly discards marks that nobody else supported.

In the record

astronomyVolunteer labels train a model to find asteroid trails in Hubble archive imagesVolunteers' click positions marking asteroid trails on Hubble image cutouts were pooled into single detections using HDBSCAN point clustering, with a minimum cluster size of five.astronomyKepler planets sorted by host star's Galactic orbit show differing eccentricitiesKepler planet hosts crossmatched with APOGEE abundances were cut to main-sequence stars and split into chemical thick- and thin-disk groups by a linear magnesium-iron boundary.astronomyNeural networks turn simulated cluster mass maps into X-ray and SZ imagesSimulated cluster halos were projected into mass, SZ and X-ray maps, then normalised by a log transform and standardisation, chosen after testing five schemes.astronomyNeural networks stand in for a slow model of interstellar gas cloudsA precomputed grid of interstellar medium model runs was split into training and test sets, with intensities and most input parameters taken in log scale and standardised.

Limits

Clustering finds groups whether or not any exist. It will happily split a smooth continuum in two, and the split it offers depends on which measurements were used, how they were rescaled and what settings were chosen, so a different but equally reasonable setup can give different groups. Astronomical measurements carry uncertainties and gaps that the method does not know about, and objects at very different distances can look similar. The groups themselves carry no physical meaning until someone argues for one, which is why sharp hand-drawn boundaries and follow-up checks remain common.

Written by AIxSci: the background from general knowledge, the examples from checked records.