Deep Learning Identifies High-z Galaxies in a Central Blue Nugget Phase in a Characteristic Mass Range

TL;DR

This paper employs machine learning, specifically CNNs, to identify the blue nugget phase in high-redshift galaxies from images, bridging simulations and observations and revealing a characteristic mass range for these galaxies.

Contribution

It introduces a CNN-based method to classify galaxy phases in both simulated and observed data, enabling direct comparison and revealing a characteristic mass range for blue nuggets.

Findings

CNN accurately identifies galaxy phases in simulations.

Observed blue nuggets are found in galaxies with stellar masses $10^{9.2-10.3} M_\u2299$.

The method facilitates comparison between theory and observations.

Abstract

We use machine learning to identify in color images of high-redshift galaxies an astrophysical phenomenon predicted by cosmological simulations. This phenomenon, called the blue nugget (BN) phase, is the compact star-forming phase in the central regions of many growing galaxies that follows an earlier phase of gas compaction and is followed by a central quenching phase. We train a Convolutional Neural Network (CNN) with mock "observed" images of simulated galaxies at three phases of evolution: pre-BN, BN and post-BN, and demonstrate that the CNN successfully retrieves the three phases in other simulated galaxies. We show that BNs are identified by the CNN within a time window of $\sim0.15$ Hubble times. When the trained CNN is applied to observed galaxies from the CANDELS survey at $z=1-3$, it successfully identifies galaxies at the three phases. We find that the observed BNs are preferentially found in galaxies at a characteristic stellar mass range, $10^{9.2-10.3} M_\odot$ at all redshifts. This is consistent with the characteristic galaxy mass for BNs as detected in the simulations, and is meaningful because it is revealed in the observations when the direct information concerning the total galaxy luminosity has been eliminated from the training set. This technique can be applied to the classification of other astrophysical phenomena for improved comparison of theory and observations in the era of large imaging surveys and cosmological simulations.