Oral Presentation

Mining for the Protoclusters at z〜4 from HSC-SSP photometric dataset with Deep Learning

Author(s): Yoshihiro Takeda, Nobunari Kashikawa, Kei Ito (U. Tokyo), Kent Fujiwara (LINE Corporation), Jun Toshikawa (U. Bath), Yongming Liang (ICRR), Rikako Ishimoto, Takehiro Yoshioka, Junya Arita (U. Tokyo), Mariko Kubo (Tohoku U.), Hisakazu Uchiyama (Ehime U.)

Presenter: Yoshihiro Takeda (The University of Tokyo)

Protoclusters are high-density regions at high-z that are expected to evolve into the clusters by z=0. They are good targets for understanding galaxy evolution which is accelerated by environmental effects. However, the identified protoclusters beyond z>3 are few due to large redshift uncertainties, preventing us from studying them. In this talk, I will present a new point-based deep learning model, PCFNet, to detect protocluster member candidates from a g-dropout catalog. We use the sky distribution, i band magnitude, color (g-i), and the estimate of the redshift probability density function of galaxies surrounding a target galaxy on the sky. A conventional model based only on the surface number density of g-dropout galaxies has a recall of 1.5±0.1% and a precision of 38±2%, while the PCFNet achieves a recall of 7.5±0.2% and a precision of 44±1%. The PCFNet is able to detect 5 times more protocluster member candidates more accurately. Moreover, the PCFNet can detect lower-mass protoclusters candidates than before, providing a wider dynamic range of halo mass to allow us to get closer to a more general picture of protoclusters. We applied the PCFNet to observational data, the HSC-SSP Deep/UltraDeep layer, and obtained 88 protocluster candidates from an effective area of about 13 deg^2. We are proposing the spectroscopic observation and plan to apply the PCFNet to wider regions such as HSC-SSP wide layer.