Candidate generation in astronomy and astrophysics

How it is used

Astronomy studies objects that cannot be handled or rebuilt: stars, planets, the gas between them, the Sun. Researchers work instead from light collected by telescopes and from models of the physics they believe is at work. Both leave gaps. There are more possible configurations of a model than anyone can test, and more hours of sky than anyone can inspect. Candidate generation uses software that proposes items rather than labelling them: the program suggests a shortlist, and the shortlist is then checked by simulation, by measurement, or by eye. In the examples here, one program proposed simplified chemical schemes for planetary atmospheres, and another produced synthetic pictures of the Sun.

Two families of method appear. The first is evolutionary search, where a possible answer is written down as a string of yes-or-no choices, many such strings are scored, and the better ones are mixed and altered to make the next batch. The second is a generative model, trained on a large set of real images until it can turn noise into a picture that resembles the training data. Neither method is told what the right answer is. Both produce material for later evaluation: a scheme to be run through a chemistry code, or an image to be compared against genuine observations.

Why researchers use it

Researchers reach for these methods when the space of possibilities is far too large to enumerate and when there is no formula that points straight to a good choice. Deciding which chemical reactions matter in a planet's atmosphere is one such problem: the full network is unwieldy, and the useful subsets are not obvious in advance. A search that tries many combinations and keeps what scores well stands in for the expert judgement that would otherwise trim the list by hand, slowly and with no record of what was rejected.

Generative models serve a different shortage. Some events in the sky are rare, so the pictures of them are few, which makes it hard to train or test any further method. A model that has learnt the look of solar images can supply extra examples on demand. Here the generated image stands in for an observation that was never made, with the understanding that it carries no new information about the Sun itself, only the patterns already present in the archive it learnt from.

In the record

astronomyGenetic algorithm trims chemical reaction lists for hot Jupiter atmosphere modelsA genetic algorithm searched for reduced chemical networks, each candidate written as a binary string whose genes mark whether a forward-and-reverse reaction pair is kept.astronomyDiffusion model generates synthetic images of the Sun sorted by flare strengthA denoising diffusion model with a U-Net backbone and self-attention layers was trained on solar ultraviolet images to generate new synthetic frames.

Limits

A proposal is not a finding. An evolutionary search returns what scored well on the chosen measure, which may reward a scheme that fits the test cases and fails elsewhere, and it gives no guarantee that better options were not missed. A generative model can only recombine what its training archive held, so it will reproduce that archive's gaps and instrument quirks while looking convincing. Synthetic images resemble observations without being evidence about the sky, and using them to train or test other tools can quietly circulate the same assumptions. Independent checking remains necessary throughout.

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