Artificial intelligence can accurately predict the age of supernovae and other rare stellar explosions within milliseconds of a telescope spotting them.
That could prove useful for projects like the Vera C. Rubin Observatory鈥檚 Legacy Survey of Space and Time, which is slated to begin observing in 2025. The 10-year survey of the southern sky will take 15 terabytes of data per night and could spot more than 10 million potential cosmic events each night. That may boost supernovae discoveries 100-fold.
鈥淚t鈥檚 going to observe potentially millions of alerts every single night, which is an order of magnitude more than any survey we鈥檝e had before,鈥� says at the Massachusetts Institute of Technology. 鈥淭hat means we really do need machine learning approaches to tackle some of these problems that previously we might have been able to tackle by human eyes.鈥�
Advertisement
Muthukrishna and his colleagues developed an automatic system for identifying the age of each event as quickly as possible once it shows up in the telescope surveys. That would enable astronomers to use other telescopes to analyse cosmic events closer to the time of explosion.
The researchers trained one machine learning algorithm to predict the time of first light 鈥� which usually coincides with the supernova explosion 鈥� and a second to predict when the event鈥檚 brightness will reach its maximum. Exploding stars often brighten very quickly over about 10 days before gradually fading. Knowing the peak time and the explosion鈥檚 age will let astronomers quickly make follow-up observations before they fade, says at the California Institute of Technology.
After training the algorithms on supernovae events previously spotted by the Zwicky Transient Facility and the Transiting Exoplanet Survey Satellite, the researchers showed that the AI could predict the time of first light with an average error of just two days and the time of maximum light with an average error of about four days. The system can make predictions for 100 objects within just 1.7 milliseconds when running on a relatively older NVIDIA chip. It could work even faster with more computational power, says Muthukrishna.
But Mahabal cautioned that the training dataset for this study is relatively small, and he questioned whether the AI-powered approach could tell supernovae apart from other bright objects such as flaring dwarf stars or energy released by supermassive black holes.
Next, the researchers plan to test how the AI performs during live observations on tens of thousands of new objects being detected nightly by the Young Supernova Experiment based in Hawaii.
arXiv
