Machine Learning for Strong Lensing: Preparing for Discovery in Next-generation Astronomical Surveys
Speaker: Margherita Grespan (NCBJ, Poland -> Oxford University)
Abstract:
The Rubin Observatory and Euclid mission observe billions of galaxies, with approximately 100,000 expected to be strong gravitational lenses (SGLs). Detecting these rare systems is important for understanding dark matter and galaxy evolution, but requires machine learning to efficiently process vast datasets.
In this seminar, I will present my work on applying transformer encoders to 221 deg² of the Kilo Degree Survey (KiDS) to search for new SGL candidates. Transformers, initially trained on simulated data, are fine-tuned with real data, reducing false positives by 70% and identifying 71 high-confidence SGLs, though precision remains below 1%.
Can we bypass training on simulations altogether? To explore this, I will discuss an active learning approach using the Astronomaly pipeline, which achieves nearly 5% precision with only a few thousand labeled objects.
For next-generation surveys, combining these two methods has the potential to significantly enhance discovery.