A stellar performance – Dr Heloise Stevance is going supernova in her field of exploding stars

Dr Heloise Stevance is revolutionising the way astronomers use AI to recognise and interpret the mechanisms behind stars going supernova. Her Virtual Research Assistant is now in operation for the ATLAS sky survey, with plans to roll this out to the Large Synoptic Survey of Space and Time.


“Being a scientist is being nosy with the Universe” says Heloise, who was drawn to science from a young age. She enjoys figuring out how things work, solving problems, and discovering the answers. It seems fitting that Heloise was drawn to the “biggest fireworks” that the laws of physics can create. Her job consists of playing ‘spot the difference’ with the Universe, finding stars that have exploded in distant galaxies. These objects are important to study, because they create some of the essential building blocks to the cosmos and life as we know it. “When I find a blob in a telescope (or when my bots do), I know that somewhere millions of years ago, a star collapsed on itself and exploded with the light of a billion suns, and when that star does weird things before or after it dies, I can’t help but ask ‘why’?” 

heloise telescope chile

Heloise next to a telescope in Chile

Heloise’s research involves making bots to help teams sift through the many thousands of credible alerts received each day in our sky surveys. To do this, she considers not only why the star is doing a particular thing, but also how the human makes a decision. To make bots work well, she must think about the scientist first and foremost, and she relishes being in a job where she can both look out to the stars and work closely with colleagues. 

Incorporating machine learning (ML) into her research was unavoidable, as sky surveys deal with huge amounts of data; for example, in the ATLAS sky survey, tens of millions of alerts are received every night. It would take a human one year – with no breaks – to sort through these! The challenge in incorporating ML was finding the right methods and tools. Heloise says: “Machine learning is both over-hyped and under-hyped; some techniques are held up on a pedestal, whereas others are rendered very obscure”. She gives the example of Large Language Models, which are considered revolutionary across the board. However, Heloise says, picking the right ML method is not about giving in to our sci-fi fantasy, but what is more relevant to real research cases. She credits her mentor Stephen Roberts for helping her identify the best ML techniques, and advises someone new to using AI and ML to find someone who has experience bringing real systems into the world, not just someone who has built a shiny proof of concept based on the latest fad. 

heloise milky way chile

Heloise took this photo of the Milky Way with her phone while observing in Chile

Heloise’s research has resulted in the creation of her ATLAS Virtual Research Assistant (VRA), which has been in operation alongside human teams for the ATLAS sky survey since August 2024, and has helped reduced workload by 85% with no loss of telescope follow-up opportunity. Having the experience of designing and creating the VRA and being able to bring it into an already established project has given Heloise the training she needed to make similar systems for the upcoming Large Synoptic Survey of Space and Time (LSST), which will be running for the next 10 years at the Vera Rubin Observatory. Heloise is delighted to not only have shared this tool with the community but also advocate that features-based machine learning methods are under leveraged in her field; they are trainable on only a few thousand samples, highly interpretable, and understand the absence of data and can derive meaning from that. Most importantly, they offer a direct way to inject expertise into models through smart feature engineering.  

Heloise credits receiving the Schmidt AI in Science Fellowship as the most defining moment in her career. It allowed her to combine her love of trying to understand the universe with her love of computing.  

Heloise is looking ahead to the new possibilities and opportunities the Vera Rubin Observatory will provide, and is building her little army of bots to help her look at the right place at the right time. In 1 year only, the survey will capture more data than all previous sky surveys in human history combined. She will be looking for the most massive stars in the Universe dying in real time, as that will help the understanding of how they shed their outer layers just before the final supernova explosion. This will help scientists to better understand how stars give matter back to the Universe in the final stage of their life.  

Machine learning is both over-hyped and under-hyped; some techniques are held up on a pedestal, whereas others are rendered very obscure