Oral Presentation
A method to find high-z protoclusters from photometric catalog of dropout galaxies by Deep Learning
Presenter: Yoshihiro Takeda (The University of Tokyo)
Protoclusters are regions of high galaxy number density where galaxy evolution is accelerated by environmental effects such as mergers and gas inflow and are a good target for understanding galaxy evolution. Toshikawa et al. (2018) succeeded in finding 179 unique protocluster candidates based on the g-dropout catalog based on the Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP). However, their method was focused only on the super-dense region due to the need for high purity. Thus, the sample size was small and there was a possibility that the detected protocluster was biased toward extreme cases such as “Coma-type” (>10^15 M_sun at z=0) progenitors. Another method is needed to get a more general picture of protocluster. In this talk, I present a new method to find proto-cluster candidates from the distribution of g-dropout galaxies using PointNet, a neural network capable of handling 3D point cloud data. We used (g-i) color as a proxy of the distance in LOS because the g-dropout catalog has no redshift information. We trained PointNet using the PCcone (Araya-araya et al. 2021) and obtained 5.7±0.9 times candidates of protocluster members than 2D density selection at 4σ. It allows us to lower the threshold for detection and identify more common lower-mass protoclusters. In the near future, we plan to apply this method to observational data such as the HSC-SSP DUD/Wide to construct a new protocluster sample.
asiaa.sinica.edu.tw