HOLISMOKES. VIII. High-redshift, strong-lens search in the Hyper Suprime-Cam Subaru Strategic Program

Abstract

We carry out a dedicated search for strong-lens systems with high-redshift lens galaxies with the goal of extending strong lensingassisted galaxy evolutionary studies to earlier cosmic time. Two strong-lens classifiers are constructed from a deep residual network and trained with datasets of different lens redshift and brightness distributions. We classify a sample of 5,356,628 pre-selected objects from the Wide layer fields in the second public data release of the Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP) by applying the two classifiers to their HSC gri-filter cutouts. Cutting off at thresholds that correspond to a false-positive rate of 10 −3 on our test set, the two classifiers identify 5,468 and 6,119 strong-lens candidates. Visually inspecting the cutouts of those candidates results in 735 grade-A/B strong-lens candidates in total, of which 277 candidates are discovered for the first time. This is the single largest set of galaxy-scale strong-lens candidates discovered with HSC data to date, and nearly half of it (331/735) contains lens galaxies with photometric redshifts above 0.6. Our discoveries will serve as a valuable target list for ongoing and scheduled spectroscopic surveys such as the Dark Energy Spectroscopic Instrument, the Subaru Prime Focus Spectrograph project, and the Maunakea Spectroscopic Explorer.

HOLISMOKES -- IX. Neural network inference of strong-lens parameters and uncertainties from ground-based images

“…Since there are already thousands of known lens candidates in HSC (e.g., Wong et al 2018;Sonnenfeld et al 2019Sonnenfeld et al , 2020Jaelani et al 2020a,b;Jaelani et al in prep. ;Cañameras et al 2021a;Shu et al 2022), and we expect hundreds of thousands more observed by LSST and Euclid (Collett 2015), CNNs would be perfectly suited to analyze this amount of data in an acceptable amount of time.…”

“…Since we gained in our lens search projects (Cañameras et al 2021b;Cañameras et al in prep. ;Shu et al 2022) better performance by applying a square-root stretching, we also tested this for our modeling network. Hence, we did no longer pass the images to the network, but rather the square-root of the images after setting all negative background pixels to zero.…”