Description
Euclid’s space-based resolution over a third of the sky makes it the perfect tool for finding strong gravitational lenses, being expected to increase the total number by two orders of magnitude. To find strong lenses in the Euclid Q1 data, I repurposed the Zoobot machine learning model that was pretrained on GalaxyZoo morphologies. In combination with 4 other machine learning networks, we searched for strong lens candidates in Q1, which we validated with human visual inspection. This resulted in our discovery of 500 strong lenses. The finetuned Zoobot model shows a substantial improvement in purity over previous lens searches, with 160 strong lenses in the top 1000 ranked images - a testament to the power of transfer learning. Although scaling up visual inspection is challenging, I will demonstrate how iterative machine learning training, where a network can learn from its mistakes, is able to further increase lens-finding performance. This is essential for finding the ~100,000 strong lenses in the full Euclid Wide Survey, which will facilitate significant advancements in our understanding of cosmography, galaxy formation, and dark matter.
