Toward Complete Merger Identification at Cosmic Noon with Deep Learning

TL;DR

This work demonstrates that a ResNet18 CNN trained on mock HST CANDELS-like images from the IllustrisTNG50 simulation can identify a broad spectrum of galaxy mergers, including minor and low-mass systems at $1<z<1.5$, achieving about 73% accuracy. The authors employ Grad-CAM and UMAP to interpret the model, revealing that the latent space correlates with stellar mass and sSFR, while identifying an upper-performance limit due to viewing-angle effects. The study highlights both the promise and current limitations of ML-based merger catalogs for cosmic noon and outlines concrete steps to improve with larger simulations and more balanced training sets. These findings support the feasibility of complete merger identification in upcoming deep imaging surveys and offer a roadmap to reduce biases in merger demographics across a wide mass and merger-stages regime.

Abstract

As we enter the era of large imaging surveys such as $\textit{Roman}$, Rubin, and $\textit{Euclid}$, a deeper understanding of potential biases and selection effects in optical astronomical catalogs created with the use of ML-based methods is paramount. This work focuses on a deeper understanding of the performance and limitations of deep learning-based classifiers as tools for galaxy merger identification. We train a ResNet18 model on mock Hubble Space Telescope CANDELS images from the IllustrisTNG50 simulation. Our focus is on a more challenging classification of galaxy mergers and nonmergers at higher redshifts $1<z<1.5$, including minor mergers and lower mass galaxies down to the stellar mass of $10^8 M_\odot$. We demonstrate, for the first time, that a deep learning model, such as the one developed in this work, can successfully identify even minor and low mass mergers even at these redshifts. Our model achieves overall accuracy, purity, and completeness of 73%. We show that some galaxy mergers can only be identified from certain observation angles, leading to a potential upper limit in overall accuracy. Using Grad-CAMs and UMAPs, we more deeply examine the performance and observe a visible gradient in the latent space with stellar mass and specific star formation rate, but no visible gradient with merger mass ratio or merger stage.

Figures (1)

• Figure 1: Top row: Box plots showing the 25th-75th percentile (inter-quartile range, IQR) of the data inside the box, with the whiskers stretching to the furthest point within 1.5 times the IQR. The remaining points are outside of that range. These show the overall distributions of galaxies by merger mass ratio (left; dotted line separates major and minor mergers) and stellar masses (center and right), which are identified correctly from a given number of observing angles. Middle row: Grad-CAM showing that the network focuses on the central galaxy when making its prediction. Bottom row: UMAPs showing the true class by shape (nonmergers triangle, mergers circle) and colored by stellar mass (left) and sSFR (right), respectively, which show a clear gradient.