When: February 12, 2018 @ 12:30 – 1:30pm
Where: PAB, WRF Data Science Studio, 6th floor
Big Data Cosmology with Subaru HSC Survey
Subaru HSC is ongoing and is planned to survey a total of ~1200 square degrees by 2019. We have already conducted 180-nights observation, and collected about 65% of the total. We use the data to detect and classify distant supernovae and to reconstruct the large-scale cosmic density field in 3D. To this end, we have developed a new “machine” adopting AUC boosting and pAUC methods, as well as commonly used Random Forest and DNN. The machine detected ~1500 supernovae including faint ones down to 26-th magnitude. We have also developed a multi-label classifier (Type Ia, Ibc, IIP, IIN, IIL) and used it successfully to extract a few tens high-redshift Type Ia supernovae, which have been sent for spectroscopic observations by HST. For cosmological parameter estimation, we have developed a fast, machine-learned “emulator” that calculates statistical quantities of weak lensing. We run 200 supercomputer simulations of cosmic structure formation and use the outputs to train and develop the emulator (effectively a python package). Cross-validation study shows that the emulator predicts the gravitational lensing effects on the matter distribution and on the clustering of galaxies with an accuracy of 3 percent. We will integrate the emulator in our Markov-Chain Monte-Carlo program to infer the main cosmological parameters such as the matter density and the density fluctuation amplitude. Finally, we are developing a CNN that can calculate basic strong lensing parameters such as Einstein radius and lens ellipticity from observed multi-band images. I discuss the future prospects for LSST.