Supernova Detection
summary
Difference imaging produces 1.5 million detections per night with only ~100 supernovae and 3–4 new young, follow-up-worthy events, so machine learning replaces manual scanning. A first classifier used ~1,000 human-labeled examples; RealBogus 2.0 trains on ~78,000, cuts features to ~42→fewer, achieves <1% false positive rates. Autoscan reduces 30 people to 2 with ~1% extra contamination and ~6% missed real events. Future surveys will produce one trillion detections (ZTF) and seven trillion (LSST); CMU-DeepLens classifies 100M candidates in nine hours on one GPU; NERSC is used to brute-force orbit searches.
AI-written summary · about
chapters (19)
- 0:00Introduction to sky changes
- 1:00History of astronomical search
- 3:29Modern supernova detection
- 5:16The bottleneck of manual scanning
- 7:28Palomar transient factory
- 10:06Machine learning automation
- 13:02Citizen science participation
- 15:48Real-time follow-up impact
- 18:58Improved automated classifiers
- 23:06Dark energy survey application
- 26:44Future survey infrastructure
- 30:31Alert-based astronomy
- 31:51Lensed supernovae potential
- 37:29Finding rare lensing events
- 38:46Transient brokers and ML
- 40:17Deep learning for lens detection
- 41:41Searching for planet nine
- 46:04Conclusion and outlook
- 46:51Q&A session
related talks
41mExploring Dark Energy & Star Formation w/ ML
37mCounting the Stars
01h 01mWe are the Guardians of our Future
26mHit-reco: ProtoDUNE Raw Data Denoising with DL models
23mHunting Exoplanets by Direct Imaging with Python
42mIn the Shadow of the Black Hole
22mAstronomy with Gravitational Waves
46mChasing the Dark Universe with Euclid and Python: Unveiling the Secrets of the Cosmos