An AI scheduler has chosen observations at Chile’s Blanco telescope and adjusted its plan as conditions changed. Reported on 31 July 2026, the SkAI project completed two observing runs that spring and summer with the telescope’s Dark Energy Camera.[1]
Learn the next observation
Researchers trained a deep-learning model on historical Dark Energy Survey observations. Given the telescope’s position, it predicted the next observation and learned from errors against the recorded choice. The team reports performance comparable to human scheduling, without a measured efficiency advantage.[1]
Different science needs different skies
The earlier Dark Energy Survey already automated exposure selection through its ObsTac software. It accounted for the date, weather, Moon and completed work. The new research concerns learning the scheduling policy from experience.[2][1]
What an observing plan must balance
| Scientific need | Scheduling consequence |
|---|---|
| Galaxy shapes | Reserve sharp imaging conditions for measurements sensitive to blur.Source 2 |
| Supernova brightness | Return to the same fields regularly to capture change.Source 2 |
| Survey coverage | Fill unfinished areas and repeat inadequate exposures.Source 2 |
A useful observation depends on the question being asked. Conditions unsuitable for measuring subtle galaxy shapes may still support a brightness measurement. An overdue supernova visit can take priority because another sharp image elsewhere will not repair the missing point in its light curve.[2]
The telescope must also hit its target
A separate South Pole Telescope study addressed the physical accuracy of pointing. Temperature differences deform the instrument’s structure. Two machine-learning models used weather and telescope sensors to estimate corrections to its horizontal direction and elevation, applied before each science-target observation.[3]
During 1,173 test observations on 7–14 April 2024, average combined pointing error was 12.2 arcseconds, compared with an estimated 15.6 arcseconds without the learned corrections: a 22% reduction. The comparison reconstructs the uncorrected errors from the pointing equations. An arcsecond is one 3,600th of a degree.[3]
The strongest improvement occurred where training data were densest. These results concern instrument alignment, separately from SkAI’s choice of observations.[3]
SkAI’s next objective is to outperform human scheduling. Its first observing runs provide a starting point for testing whether learned choices can improve the scientific return from telescope time.[1]
Sources & further reading
- First AI-driven telescope goes stargazing Northwestern University · 31 Jul 2026 · Accessed 12 Oct 2026
- Survey Strategy Dark Energy Survey · Accessed 12 Oct 2026
- Pointing Accuracy Improvements for the South Pole Telescope with Machine Learning Chichura and colleagues, arXiv · 30 Oct 2025 · Accessed 12 Oct 2026