Astronomers and their machines
This article by Dr Heloise Stevance is extracted from the current issue of Observatory magazine
The history of modern astronomy is one of increased partnership between humans and machines. From Galileo mapping the moons of Mars with a refracting telescope in the 17th century, to the late-20th-century invention of digital detectors and electronic computers which can detect the explosions of stars millions of light-years away, successive improvements in technology have allowed astronomers to capture more data, better data, and faster than ever.
Over the last decade, our relationship with our machines has once again been transformed. Whether you call it statistical learning, machine learning or artificial intelligence, new methods and increasingly user-friendly software have allowed our computers to be trained on our wealth of data and learn to recognise complex patterns. In addition to helping us gather, store and process data, our machines can now help us interpret them. In sky surveys like ATLAS (Asteroid Terrestrial impact Last Alert System), which look for near-earth asteroids and new cosmic explosions, each night brings a new game of ‘spot the difference’. But instead of looking at 10 differences, humans are bombarded with 10 million. Most of these are artefacts (bogus detections), much of the rest are known variable stars, and a few (roughly 100 a week) are the shiny deaths of distant stars.
Finding these needles in the cosmic haystack requires a mixture of very simple techniques (discarding noisy images) and very complex ones (computer vision) trained on large samples of past data to distinguish the good from the bad. At each step of automation, a key ingredient to maintain peace between astronomers and their machines is trust. Researchers need to know that the tools they use are reliable, stable and that their behavior is understood (if not by them, by peers who developed and tested the methods). Reproducibility is an inherent aspect of scientific endeavour, and we cannot allow this trust between scientists and their machines to erode.
If the 2010s gave us the Deep Learning Revolution and computer vision, the 2020s are the decade of Large Language Models, with computers that not only recognise your face in family photos but also talk to you as if they were breathing, thinking human beings. For astronomy and science in general, these new machines are different from previous leaps in technology. Not just for their potential, but for their danger. The scale of computing power required to thoroughly test generative AI models means that this is inaccessible to most academic teams. What does that mean for our trust in our machines? A 2024 report by the Royal Society (‘Science in the Age of AI’) summarises the opportunities and challenges ahead.
In 2025, the Vera Rubin Observatory in Chile will begin operations. It will allow us to probe the nature of dark matter, catalogue asteroids and comets, map our own galaxy like never before, and increase our discovery rate of cosmic explosions by a factor of 100 to 1,000. As a computational astronomer developing virtual assistants for sky surveys, I am acutely aware that fostering our relationship with our machines has never been more important. We need to be passionate about the potential that new machine learning tools bring, yet also profoundly sceptical of their applicability for science. We must preserve reproducibility above all, as the data we gather and process now will affect generations of astronomers to come. We must pick our tools because they are right for the job, because they can be thoroughly tested. Finally, we must strengthen our relationships to other humans: the statisticians, the computer scientists.
The future of astronomy is bright, and it will require more than a connection to our machines.
Dr Heloise Stevance is one of the first Eric and Wendy Schmidt AI in Science Fellows in the Sub-Department of Astrophysics, University of Oxford. Heloise joined as one of seven new Research Associates focused on astrophysics at college.
Read Observatory magazine 2025
