AI in astronomy and astrophysics

How AI is used

Astronomy is the study of objects beyond the Earth, from single stars to the whole distribution of matter in the universe. Most of its data arrives as images or as spectra, which are measurements of how much light an object gives off at each colour. Modern telescopes produce far more of both than anyone can inspect. AI is used here mainly for two jobs. The first is triage: deciding which of the many flagged changes in the sky are real, interesting objects rather than noise or already-known variable sources. The second is measurement: turning raw light into numbers such as a star's temperature, its chemical make-up, or the mass of a cluster of galaxies.

Models in this field are usually trained on one of two things. Some are trained on labelled real observations, where a smaller, more careful survey has already supplied trusted answers for a subset of objects, and the model learns to reproduce those answers from cheaper data. Others are trained on simulations, because physics can generate as many synthetic examples as needed, complete with the true values that the real sky never reveals. Both routes appear in the examples here, and some studies combine simulated and real data in the same pipeline.

Why researchers use it

Researchers reach for these methods because the connection between what a telescope records and what is physically there is complicated, indirect and often not expressible as a tidy formula. The light from a star is shaped by its temperature, its gravity and every element in its atmosphere at once. The shapes of distant galaxies are distorted slightly by all the mass lying between them and us. Working backwards from the observation to the cause is the hard part, and a trained model offers a way to do it without writing that inversion down by hand.

The other reason is scale and speed. Surveys that scan the sky repeatedly throw up changes faster than people can look at them, and some of those changes fade within days, so a decision has to be made quickly enough to point another telescope at the object while it is still there. AI also stands in for human attention in a narrower sense: it replaces the work of experienced observers scanning candidate lists, and encodes the judgements they used to make about what is worth following up.

In the record

Limits

These models are only as trustworthy as the examples they learn from. Where training rests on simulations, the model inherits whatever the simulation assumes, and the real sky may differ in ways nobody noticed. Where it rests on labelled observations, it works best within the range of objects that were labelled, and may fail quietly on anything rarer or stranger, which is often what astronomers most want to find. The output is also usually a number without an explanation, so checking it against physical reasoning takes separate work.

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

What the AI was for

• Classification
• Property prediction
• Simulation surrogate
• Detection
• Denoising
• Structure determination
• Anomaly detection
• Experimental design
• Candidate generation
• Segmentation

Model families

• Convolutional neural networks
• Multilayer perceptrons
• Random forests
• Autoencoders
• Clustering
• Transformers
• Gaussian process
• Linear models
• Gradient-boosted trees
• Recurrent neural networks
• Support vector machines
• Normalising flows
• Probabilistic graphical models
• Graph neural networks
• Diffusion models
• Symbolic regression

Named models

• XGBoost
• BTSbot
• braai
• Zoobot

Every article