fields/astronomy/model families
Normalising flows in astronomy and astrophysics
How astronomers use normalising flows, a kind of learned, reversible mapping that turns a simple spread of random numbers into a complicated one, and why they reach for them.
How it is used
Astronomy and astrophysics study the matter and light beyond the Earth: stars, galaxies, the gas and dark matter around them, and the structure of the universe as a whole. Almost everything is known only through patterns of brightness and motion, and almost every question is about a distribution: how stars are spread through a galaxy and how fast they move, how galaxy shapes vary, how many clusters of a given mass exist. A normalising flow learns such a distribution from examples. It is a chain of reversible mathematical steps that stretches and folds a simple starting spread, usually a plain bell curve, until it matches the data.
Because every step is reversible and smooth, the flow can be used in two directions. Run forwards, it produces new samples that look like the training data. Run backwards, it reports how likely any particular point is, and it can be differentiated, meaning the slope of the fitted distribution is available anywhere. In the research on this site, flows have been fitted to the positions and velocities of simulated galaxy stars, and combined with other learned models that compress galaxy images into a short list of numbers, so that realistic images can then be generated from a chosen set of structural measurements.
Why researchers use it
The usual alternative is to assume a shape. Astronomers have long written down tidy formulae for how light falls off across a galaxy or how stars are distributed, then fitted a few parameters. Real systems rarely obey those formulae, and bars, spiral arms, mergers and warps all break them. Histograms and smoothing kernels avoid assuming a shape but cope badly when each object carries many measured quantities at once, such as three positions and three velocities together.
A flow stands in for the assumed formula. It gives a continuous, differentiable description of a many-dimensional distribution without being told its form in advance, which matters when the next step needs derivatives: the gravitational pull in a galaxy, for instance, follows from how the distribution of stars changes from place to place. Flows also serve as fast stand-ins for slow simulation pipelines, generating plausible objects or measurements in place of rerunning the full machinery for every trial.
In the record
Limits
A flow only knows the data it was shown. Where observations are sparse, faint objects are missing, or the survey covers a limited patch of sky, the fitted distribution can invent smooth structure that is not there, and its derivatives go wrong first. Training on simulations imports the simulation's assumptions. The fit offers no interpretable parameters, so it is hard to say which physical effect produced a feature, and little warns the user when a sample or a density lies outside the range the flow actually learned.
Written by AIxSci: the background from general knowledge, the examples from checked records.