YOLO-CL cluster detection in the Rubin/LSST DC2 simulations

The next generation large ground-based telescopes like the Vera Rubin Telescope Legacy Survey of Space and Time (LSST) and space missions like Euclid and the Nancy Roman Space Telescope will deliver wide area imaging surveys at unprecedented depth. In particular, LSST will provide galaxy cluster catalogs up to z ∼ 1 that can be used to constrain cosmological models once their selection function is well-understood. Machine learning based cluster detection algorithms can be applied directly on images to circumvent systematics due to models and photometric and photometric redshift catalogs. In this work, we have applied the deep convolutional network YOLO for CLuster detection (YOLO-CL) to LSST simulations from the Dark Energy Science Collaboration Data Challenge 2 (DC2), and characterized the LSST YOLO-CL cluster selection function. We have trained and validated the network on images from a hybrid sample of (1) clusters observed in the Sloan Digital Sky Survey and detected with the red-sequence Matched-filter Probabilistic Percolation, and (2) dark matter halos with masses M200c > 1014 M⊙ from the DC2 simulation, resampled to the SDSS resolution. We quantify the completeness and purity of the YOLO-CL cluster catalog with respect to DC2 halos with M200c > 1014 M⊙. The YOLO-CL cluster catalog is 100% and 94% complete for halo mass M200c > 1014.6 M⊙ at 0.2 1014 M⊙ and redshift z ≲ 1, respectively, with only 6% false positive detections. We find that all the false positive detections are dark matter halos with 1013.4 M⊙ ≲ M200c ≲ 1014 M⊙, which corresponds to galaxy groups. We also found that the YOLO-CL selection function is almost flat with respect to the halo mass at 0.2 ≲ z ≲ 0.9. The overall performance of YOLO-CL is comparable or better than other cluster detection methods used for current and future optical and infrared surveys. YOLO-CL shows better completeness for low mass clusters when compared to current detections based on Matched Filter cluster finding algorithms applied to Stage 3 surveys using the Sunyaev Zel’dovich effect, such as SPT-3G, and detects clusters at higher redshifts than X-ray-based catalogs. Future complementary cluster catalogs detected with the Sunyaev Zel’dovich effect will reach similar mass depth and will be directly comparable with optical cluster detections in LSST, providing cluster catalogs with unprecedented coverage in area, redshift and cluster properties. The strong advantage of YOLO-CL over traditional galaxy cluster detection techniques is that it works directly on images and does not require photometric and photometric redshift catalogs, nor does it need to mask stellar sources and artifacts.