New AI method sharpens view of the Universe from supernova light

Imperial-linked research could unlock the full potential of next-generation sky surveys

A new artificial intelligence (AI) method could significantly improve how astronomers use stellar explosions to measure how the Universe expands over time.

Published in Nature Astronomy, the study introduces a powerful new approach to analysing Type Ia supernovae, a class of exploding stars that play a central role in measuring cosmic distances.

The method was developed by researchers at the International School for Advanced Studies (SISSA) in Trieste, and Imperial College London alongside collaborators at the University of Barcelona.

Making sense of supernova light

Type Ia supernovae are among the most valuable phenomena in cosmology because they act as “standardisable candles”, allowing astronomers to track how the Universe has expanded by comparing how bright they appear. These measurements provide a key tool for mapping the evolution of the Universe.

Interpreting the signals from these supernovae is far from straightforward. The observed light is influenced by a range of factors, including the properties of the progenitor star, the surrounding environment and the host galaxy in which the explosion occurs.

Traditionally, these effects are disentangled using detailed spectroscopic observations, which measure not only how much light arrives, but also how that light is distributed across different wavelengths.

However, obtaining detailed spectra for large samples of supernovae is far more time-consuming than simply collecting their brightness.

Co-author Professor Roberto Trotta, Imperial College London and SISSA, said “Collecting detailed, homogeneous spectra at multiple epochs for very large samples will be impossible, given the sheer volume of data expected in the coming years.”

For this reason, it will become increasingly important to extract reliable information even from photometric data alone – that is, from the brightness of supernovae.”

The new method, known as CIGaRS (Combined Inference and Galaxy-Related Standardisation), uses artificial intelligence and neural networks to model these effects together within a single framework. The approach uses simulations and machine learning to link observed data to the underlying physical properties of the Universe.

Rather than applying corrections step-by-step, it accounts simultaneously for factors such as galaxy properties, dust and the physical processes behind the explosions, as well as galaxy evolution.

This allows, for the first time, supernovae properties, their galactic environments and cosmological parameters to be analysed together within a single, self-consistent model.

The method can also estimate distances to the supernovae using brightness measurements alone, achieving a level of precision four time better than traditional methods.

Preparing for next-generation surveys

To test the approach, the researchers first built a catalogue containing 1,578 selected supernovae, designed to be representative of the size of current supernovae catalogues. They then extended this to a catalogue roughly ten times larger, with nearly 16,000 objects, resembling what might be collected in a single month in future observations.

The method was able to recover multiple interconnected properties at once, including how the Universe expands and how supernovae are influenced by their galactic environments.  The precision on cosmological parameters achieved using data based on images alone is comparable to that obtained in previous analyses based on smaller samples of spectroscopic measurements.  

The need for such methods is growing rapidly. Next-generation observatories such as the Vera C. Rubin Observatory are expected to discover hundreds of thousands of Type Ia supernovae each year, generating datasets far larger than those currently available.

By enabling scientists to extract more information from photometric data alone, the researchers say the method could help unlock the full potential of these surveys, making it possible to use far more of the available data, rather than relying only on the small fraction of supernovae, sometimes as little as one per cent, for which detailed spectroscopic measurements can be obtained.