April 12 2023 at 12:00 CET Giuseppe Angora from University of Ferrara.
Title: Deep Learning based search for galaxy scale-lenses in galaxy cluster environment
Abstract: In the current era of big data, the development of methods able to autonomously extract information from vast multi-dimensional datasets is playing a pivotal role. In this talk, I will present the capabilities of Convolutional Neural Networks (CNNs) to select a complete set of cluster members and to identify some of them as galaxy-galaxy strong-lenses (GGSLs). By studying these systems, we can characterize the sub-halo component of the mass distribution of galaxy clusters and test the ΛCDM structure formation paradigm. Although CNNs have been frequently used to identify GGSL in the field, I will present how these cutting-edge algorithms can be tuned to detect specifically such systems in the dense environment of galaxy clusters. These networks have been trained with simulated GGSL, where sources have been injected in real HST image cutouts exploiting high-precision cluster-lens models of eight clusters from the CLASH and HFF surveys. The simulation process is completely driven by observations, preserves the complexity of real data, and produces simulated events indistinguishable from the real ones. The best CNNs achieve a high purity-completeness level (88%-93%). This methodology is being applied to search for GGSLs around 6000 cluster members in 50 CLASH, HFF, and RELICS clusters, and can be easily extended to upcoming generations of wide-area facilities, such as VST surveys, Rubin/LSST, and Euclid.
Abstract: In the current era of big data, the development of methods able to autonomously extract information from vast multi-dimensional datasets is playing a pivotal role. In this talk, I will present the capabilities of Convolutional Neural Networks (CNNs) to select a complete set of cluster members and to identify some of them as galaxy-galaxy strong-lenses (GGSLs). By studying these systems, we can characterize the sub-halo component of the mass distribution of galaxy clusters and test the ΛCDM structure formation paradigm. Although CNNs have been frequently used to identify GGSL in the field, I will present how these cutting-edge algorithms can be tuned to detect specifically such systems in the dense environment of galaxy clusters. These networks have been trained with simulated GGSL, where sources have been injected in real HST image cutouts exploiting high-precision cluster-lens models of eight clusters from the CLASH and HFF surveys. The simulation process is completely driven by observations, preserves the complexity of real data, and produces simulated events indistinguishable from the real ones. The best CNNs achieve a high purity-completeness level (88%-93%). This methodology is being applied to search for GGSLs around 6000 cluster members in 50 CLASH, HFF, and RELICS clusters, and can be easily extended to upcoming generations of wide-area facilities, such as VST surveys, Rubin/LSST, and Euclid.