Extremely Metal-poor Representatives Explored by the Subaru Survey (EMPRESS). I. A Successful Machine-learning Selection of Metal-poor Galaxies and the Discovery of a Galaxy with M*<10⁶M and 0.016Z

We have initiated a new survey for local extremely metal-poor galaxies (EMPGs) with Subaru/Hyper Suprime-Cam (HSC) large-area (~500 deg2) optical images reaching a 5s limit of ~26 mag, about 100 times deeper than the Sloan Digital Sky Survey (SDSS). To select Z/Z o? < 0.1 EMPGs from ~40 million sources detected in the Subaru images, we first develop a machine-learning (ML) classifier based on a deep neural network algorithm with a training data set consisting of optical photometry of galaxy, star, and QSO models. We test our ML classifier with SDSS objects having spectroscopic metallicity measurements and confirm that our ML classifier accomplishes 86% completeness and 46% purity EMPG classifications with photometric data. Applying our ML classifier to the photometric data of the Subaru sources, as well as faint SDSS objects with no spectroscopic data, we obtain 27 and 86 EMPG candidates from the Subaru and SDSS photometric data, respectively. We conduct optical follow-up spectroscopy for 10 of our EMPG candidates with Magellan/LDSS-3+MagE, Keck/DEIMOS, and Subaru/FOCAS and find that the 10 EMPG candidates are star-forming galaxies at z = 0.007-0.03 with large Hß equivalent widths of 104-265 Å, stellar masses of log() = 5.0-7.1, and high specific star formation rates of ~300 Gyr-1, which are similar to those of early galaxies at z ? 6 reported recently. We spectroscopically confirm that 3 out of 10 candidates are truly EMPGs with Z/Z o? < 0.1, one of which is HSC J1631+4426, the most metal-poor galaxy, with Z/Z o? = 0.016, ever reported.