Searching for strong lensing by late-type galaxies in UNIONS

TL;DR

The study targets strong lensing by late-type, edge-on galaxies in the UNIONS CFIS data to overcome biases toward massive early-type deflectors. It combines a data-driven mock-lens training pipeline with a convolutional neural network (CMU DeepLens) and extensive visual inspection to identify edge-on lens candidates, while also estimating the prevalence of such lenses via a prevalence study and Bayesian extrapolation. The authors report the discovery of 4 grade A, 20 grade B, and 58 grade C edge-on candidates (plus non-edge-on lenses), validate two candidates spectroscopically, and discuss limitations from blending and domain adaptation, arguing that higher-resolution surveys like Euclid will be decisive for confirming and expanding this rare population. The work demonstrates feasibility for AI-guided edge-on lens searches in wide-area optical surveys, highlights the importance of empirical validation over simulated metrics, and outlines a practical path toward larger, robust samples in the coming years.

Abstract

Recent wide-field galaxy surveys have led to an explosion in the number of galaxy-scale strong gravitational lens candidates. However, the vast majority of them feature massive luminous red galaxies as the main deflectors, with late-type galaxies being vastly under-represented. This work presents a dedicated search for lensing by edge-on late-type galaxies in the Ultraviolet Near Infrared Optical Northern Survey (UNIONS). The search covers $3600$ deg$^2$ of $r$-band observations taken from the Canada-France-Hawaii Telescope. We considered all sources with magnitudes in the range $17 < r < 20.5$, without any colour pre-selection, yielding a parent sample of seven million sources. We characterised our parent sample via the visual inspection of $120\,000$ sources selected at random. From it, we estimate, with a 68% confidence interval, that 1 in every $30\,000$ sources is an edge-on lens candidate, with at least eight high-quality candidates in the parent sample. Our search relied on a convolutional neural network (CNN) to select a reduced sample of candidates, which we followed with a visual inspection to curate the final sample. The CNN was trained from scratch using simulated $r$-band observations of edge-on lenses, and real observations of non-lenses. We found 61 good edge-on lens candidates using the CNN. Moreover, combining the CNN candidates with those found serendipitously and those identified while characterising the parent sample, we discovered 4 grade A, 20 grade B, and 58 grade C edge-on lens candidates, effectively doubling the known sample of these systems. We also discovered 16 grade A, 16 grade B, and 18 grade C lens candidates of other types. Finally, based on the characterisation of the parent sample, we estimate that our search found around 60% of the bright grade A and B edge-on lens candidates within the parent sample.

Figures (12)

• Figure 1: Visualisation tools used for the visual inspections. Left panel: Mosaic tool showing an $8\times8$ grid whilst highlighting sources classified as 'lenses' with a capitalised 'L' instead of the UNIONS stamp. Right panel: One-by-one sequential tool showing a spiral galaxy on the left side and the larger field of view from Legacy Survey on the right side. The source is classified as a 'non-lens' and further sub-classified as a 'spiral'. Both applications are using logarithmic scale and a grey colour-map for the UNIONS stamps.
• Figure 2: Lens candidates discovered during the prevalence study sorted by visual inspection grade and right ascension. We show a colour composite of the $r$- and $u$-bands when the $u$-band data are available, otherwise we just display a grey scale $r$-band image. The number in the bottom left is the CNN score, and the letter in the bottom right corresponds to the final human classification. We highlight the edge-on lenses with their name in cyan, and the candidates previously reported in SLED with an overlaid orange square.
• Figure 3: Schematic description of our lens simulation procedure. To produce each simulated stamp, we started with a potential deflector and a target Einstein radius, $\theta_{\mathrm{E}}$. Then, we picked a set of parameters for a mass model consistent with the target $\theta_{\mathrm{E}}$ and the deflector's light, and we performed a check where we made sure that $\log M$, $f_\ast$, and $\Upsilon$ are within physical ranges. After that, we picked a source using the target $z_s$ provided by the mass model parameters, and we placed it randomly around the caustics before simulating its lensing by the mass model. Next, we convolved the lensed source image with the deflector's PSF, downsampled to match the pixel size of CFIS, and added Poissonian noise. At this point, we added the original stamp of the deflector to get the final mock. Lastly, we performed a contrast check to make sure that the lensing features are visible. If it passed the check, we accepted the mock and proceeded to the next deflector. Otherwise, we repositioned the source relative to the caustics and tried a second time.
• Figure 4: Validation metrics for our training of CMU Deeplens. All the datasets are balanced. Left: Minimisation of the loss function during training. The dotted line marks the epoch with the best validation loss. Right: ROC curves for the selected model. The area under the curve is 0.9993, 0.9992, and 0.9992 for the training, validation, and test datasets, respectively. The '$\times$' marks the values at the classification threshold 0.9.
• Figure 5: Left: Histogram in logarithmic scale of the distribution of network scores for the full parent sample. Only 17514.0 of the sources had a score greater than 0.9. Right: Precision and recall versus network score for the test dataset. The dashed line marks the classification threshold applied on the parent sample, 0.9. We balanced the precision, recall, and number of sources selected to settle on a classification threshold.