Euclid: Searches for strong gravitational lenses using convolutional neural nets in Early Release Observations of the Perseus field

TL;DR

This paper evaluates the effectiveness of convolutional neural networks in detecting strong gravitational lenses in Euclid survey data, highlighting current limitations and the need for improved filtering to reduce false positives.

Contribution

It provides a quantified assessment of CNN-based lens detection purity and completeness on Euclid data, revealing current challenges and potential pathways for enhancement.

Findings

CNNs perform well on simulated data but have low purity (11%) on real Euclid images.

Most false positives are identifiable as non-lenses by humans, indicating room for improved filtering.

Human classification remains essential for verifying CNN-selected lens candidates.

Abstract

The Euclid Wide Survey (EWS) is predicted to find approximately 170 000 galaxy-galaxy strong lenses from its lifetime observation of 14 000 deg^2 of the sky. Detecting this many lenses by visual inspection with professional astronomers and citizen scientists alone is infeasible. Machine learning algorithms, particularly convolutional neural networks (CNNs), have been used as an automated method of detecting strong lenses, and have proven fruitful in finding galaxy-galaxy strong lens candidates. We identify the major challenge to be the automatic detection of galaxy-galaxy strong lenses while simultaneously maintaining a low false positive rate. One aim of this research is to have a quantified starting point on the achieved purity and completeness with our current version of CNN-based detection pipelines for the VIS images of EWS. We select all sources with VIS IE < 23 mag from the Euclid Early Release Observation imaging of the Perseus field. We apply a range of CNN architectures to detect strong lenses in these cutouts. All our networks perform extremely well on simulated data sets and their respective validation sets. However, when applied to real Euclid imaging, the highest lens purity is just 11%. Among all our networks, the false positives are typically identifiable by human volunteers as, for example, spiral galaxies, multiple sources, and artefacts, implying that improvements are still possible, perhaps via a second, more interpretable lens selection filtering stage. There is currently no alternative to human classification of CNN-selected lens candidates. Given the expected 10^5 lensing systems in Euclid, this implies 10^6 objects for human classification, which while very large is not in principle intractable and not without precedent.