~/aixsci
weekly updates

fields/astronomy/what the AI was for

Attribution in astronomy and astrophysics

How astronomers use artificial intelligence to work out what produced a signal, and which parts of the data drove a model's answer, rather than simply reading off a number.

How it is used

Astronomy and astrophysics study objects that cannot be touched, moved or repeated. Everything arrives as light, or as radio waves, gathered by a telescope and turned into a brightness measurement that changes over time, a spectrum or an image. Attribution is the question of what accounts for what is seen. A dip in a star's brightness might be a planet passing in front of it, or a companion star, or a fault in the instrument. A short radio flash might be shaped by the source itself or by the gas it travelled through. Machine learning models, mostly neural networks, networks of simple numerical units trained on examples to map an input to an output, are used to apportion an observation among these possible causes.

The same models are also asked a second, internal question: which part of the input led them to their answer. Here the data are deliberately split into separate strands, one for each physical effect, so a model's reliance on each strand can be examined. In the examples on this site that means feeding a network the brightness curve, the apparent shift in the star's position, the signal half an orbit later and the alternate transits as distinct views; or feeding it X-ray light divided into soft, medium and hard energy bands alongside the motions of a cluster's galaxies. Training sets are sometimes built by simulation, so the true cause of every example is known in advance.

Why researchers use it

The traditional route to attribution is a physical model fitted to the data, with each candidate cause tested in turn. That works when the physics is tractable and the data are clean, and it struggles when neither holds. Surveys now produce far more candidate signals than can be examined by hand, and many effects, instrument noise, gas along the line of sight, hot gas and dark matter inside a galaxy cluster, overlap in ways no closed equation separates neatly. Learned models stand in for that case-by-case human judgement, and for the simplifying assumptions a fitted model has to make, such as treating a cluster as settled and symmetric when it is neither.

Attribution is also what makes an automated answer usable. A classification of a planet candidate, or a cluster mass, enters catalogues that other work depends on, so astronomers need to know whether the model reached it for the right reason or for an artefact of the instrument. Structuring the input by physical effect, and testing what the model draws on from each, lets a result be checked against the diagnostic tests astronomers already trust. Where real examples are too few or too lopsided to cover the range of possibilities, simulated data provide cases whose causes are known, so a model's apportionment can be compared with the truth.

In the record

Limits

A learned model can reach the right answer from the wrong thing: a correlation in the training data, a quirk of one instrument, or the way the input was prepared. Showing which strand of data it leaned on narrows that risk without removing it, because the strands themselves are not fully independent. Models trained on simulations inherit whatever the simulation assumes, including which causes exist at all, so an unanticipated cause may simply be sorted into a familiar box. Nor does reliance on an input establish a physical mechanism; it describes the model, not the sky.

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

Related

Every article