Autoencoders in astronomy and astrophysics

How it is used

Astronomy produces measurements in bulk: images of the Sun and of galaxies, light curves that record how bright an object is over time, radio signals spread across frequency, and strain readings from gravitational wave detectors. An autoencoder is a pair of networks. The encoder turns one such measurement into a short list of numbers, often called a latent code or representation. The decoder tries to rebuild the original from that code alone. Nothing but the data is needed to train it, because the target is the input itself. If the rebuilt version is close, the short code has kept what matters and dropped the rest.

In the examples here, the squeezed code is put to work in several ways. It can stand in for a raw signal when sorting objects into shapes or classes. It can be the starting state for a model of how a source changes with time, with a decoder that unrolls that state forward and a further network that reads physical quantities off it. Encoder and decoder layers also sit inside image generators, which learn to remove noise step by step and so can produce synthetic pictures. In every case the data are first put on a common footing: trimmed, re-binned onto a fixed grid, cleaned of interference and rescaled.

Why researchers use it

Astronomers reach for autoencoders because their data are large, repetitive and awkward to describe by hand. A single observation may hold thousands of pixels or time samples, while the thing of interest, the shape of a burst or the mass of a black hole, is captured by a handful of quantities. Choosing those quantities by eye is slow and depends on what someone already expects to see. An autoencoder learns a compact description from the data, without labels, and so can be trained on simulations or on archives of unlabelled observations.

The short code then stands in for the raw measurement in later steps: as input to a classifier, as the state of a model of change over time, or as a way of grouping observations that resemble each other. It also stands in for a rule that nobody has written down, the rule saying which parts of a noisy signal carry information and which parts are instrument or sky. Because the same encoder can take simulated and real data, work done with plentiful synthetic examples can be carried over to scarcer observations.

In the record

astronomyTwo unsupervised methods compress fast radio burst spectra into a handful of numbersLearned compact representations of radio burst dynamic spectra, using simulated spectra spanning five morphological categories and real telescope voltage data standardised onto a uniform time-frequency grid.astronomyNeural network reads simulated quasar light curves to estimate black hole propertiesA recurrent encoder compressed simulated quasar light curves into a latent state, a stochastic differential equation decoder evolved it, and a further network estimated physical parameters.astronomyDiffusion model generates synthetic images of the Sun sorted by flare strengthA diffusion model with a U-Net backbone, downsampling then upsampling with attention between, was trained on solar images to generate synthetic ones.astronomyMachine learning predicts whether a dying massive star explodes, from its density profileStellar density profiles were truncated in mass, interpolated onto a uniform grid and normalised, then learned against explode-or-not labels from two-dimensional simulations.

Limits

An autoencoder is rewarded only for rebuilding its input, so it may keep whatever varies most, including instrument artefacts and sky interference, and discard a faint feature that matters physically. The short code has no agreed meaning: its numbers are not masses or temperatures, and reading physics off them needs a further, separately checked step. Models trained on simulated data inherit whatever the simulation assumes, and may behave oddly on real observations that differ. Heavy preprocessing decides much of the outcome before learning starts. Reconstruction quality alone does not show that anything useful was learned.

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