Astronomers working in Chile have put into operation a self-driving telescope capable of learning from its own observations to determine where to direct itself next. The AI system manages telescope time without human involvement, modifying its decisions in real time based on changes in weather and light conditions.
How the AI Learns to Stargaze
The AI system was trained using years of data from the Dark Energy Survey. Scientists provided the system with historical telescope positions and asked it to forecast the next observation. Whenever its forecast was incorrect, the system would recalibrate. Through many cycles of this process, it became capable of scheduling observations without being explicitly informed about factors such as the brightness of the moon or the presence of clouds.
The AI tool was developed by researchers from Northwestern University, the University of Chicago, and Fermilab. It is currently being used at the 4-meter Blanco Telescope in Chile. This telescope, equipped with a 570-megapixel Dark Energy Camera, is among the most productive in the world. The AI system didn’t just create an observing plan but also modified it during observations as environmental conditions changed. Alex Drlica-Wagner, one of the co-leaders of the project, noted that the next objective is to teach the system to surpass human astronomers in planning observations.
Large telescopes are rare resources that astronomers from around the world seek to use. A more efficient scheduling system allows better use of the limited time available. Choosing the wrong targets could lead to unworkable data, and missed opportunities might not be recoverable for months.
Drlica-Wagner, a scientist at Fermilab and a professor at the University of Chicago, emphasized that large telescopes represent national or international assets. If everyone could use their time more efficiently, the scientific community would be able to conduct more research.
Drlica-Wagner combined his knowledge in large astronomical surveys with Vijayaraghavan’s expertise in machine learning. With their teams at the SkAI, they developed a scheduling system based on deep learning. Instead of programming the AI with rules that astronomers have developed over decades, the researchers allowed the system to learn independently.
Drlica-Wagner explained how the training process worked. The researchers showed the model where the telescope was pointing at one moment and asked it to predict the next observation. After repeating this process many times, the model learned how to schedule observations without being explicitly taught how the brightness of the moon, atmospheric conditions, or other factors impact the quality of astronomical observations.
The AI system represents a major milestone in the development of more autonomous observatories. The team successfully deployed the system on a national observatory, setting up all the necessary infrastructure. Drlica-Wagner stated that currently, the system’s performance is comparable to that of a human. He expressed confidence that in the future, the system could surpass human capabilities in scheduling observations.
Paul Chichura, a postdoctoral associate at SkAI; Rachel Hur, a PhD student at the University of Chicago; and Guillermo Damke, an associate scientist at NSF’s NOIRLab, all played key roles in deploying the system at the Blanco telescope.
A poorly positioned telescope could produce images that are too blurry or washed out by moonlight, making faint or distant objects even harder to detect.
This project, supported by the National Science Foundation (NSF)-Simons Foundation AI Institute for the Sky (SkAI), demonstrates the potential of AI to revolutionize the way telescopes are operated. The success of this project paves the way for more autonomous observatories in the future.