Beyond the Brightest: A Deep Learning Approach to Identifying Major and Minor Galaxy Mergers in CANDELS at $z \sim 1$

TL;DR

This paper presents a deep learning pipeline trained on realistic, radiative-transfer–processed mock HST CANDELS images from the IllustrisTNG50 simulation to identify galaxy mergers at $z\sim1$ across a broad stellar-mass range down to $M_*\sim10^8\,M_\odot$ and mass ratios $\mu>1:10$. Using a ResNet18 backbone with transfer learning and six viewing angles, the model achieves ~73% accuracy and AUC ~0.80, demonstrating capability to detect both major and minor mergers and both early and late stages, while also highlighting intrinsic limits due to viewing angle and star formation. The work emphasizes the need for mass- and SFR-balanced training sets and shows how interpretability tools like Grad-CAM and UMAP connect network decisions to physical galaxy properties, informing future application to JWST, Rubin, Roman, and Euclid data. It also points toward domain adaptation and combining simulations across box sizes to improve generalization to real observations. Overall, the study provides a practical, scalable approach for assembling merger catalogs in large surveys and sets the stage for extending merger studies to lower masses and higher redshifts.

Abstract

Galaxy mergers play an important role in many aspects of galaxy evolution, therefore, more accurate merger identifications are paramount for achieving a complete understanding of galaxy evolution. As we enter the era of very large imaging surveys, we are able to observe mergers extending to even lower masses and higher redshifts. Despite low-mass galaxies being more common, many previous merger identification methods were mostly calibrated for high-mass, local galaxies, which are easier to identify. To prepare for upcoming surveys, we train a convolutional neural network (CNN) using mock $\textit{HST}$ CANDELS images at $z \sim 1$ created from the IllustrisTNG50 cosmological simulation. We successfully identify galaxy mergers between a wide range of galaxies ($10^8M_\odot < M_\star < 10^{12.5}M_\odot$, and $μ>1:10$), achieving overall accuracy, purity, and completeness of $\sim73\%$. We show, for the first time, that a CNN trained on this diverse set of galaxies is capable of identifying both major and minor mergers, early and late stage mergers, as well as nonmerging galaxies, similar to that of networks trained at lower redshifts and/or higher masses (with accuracies ranging between $66-80\%$) for the first time. We discuss the inherent limits of galaxy merger identification due to orientation angle and explore the confounding variables, such as star formation, to consider when applying to real data. This network enables the exploration of the impact of previously overlooked mergers of high mass ratio and low stellar masses on galaxy evolution in CANDELS, and can be expanded to surveys from $\textit{JWST}$, Rubin, $\textit{Roman}$, and $\textit{Euclid}$.

Figures (11)

• Figure 1: Left: Distributions of merging and nonmerging galaxies' stellar masses. Center: Stacked histogram of merger mass ratios in the merging sample color coded by merger stage as defined in Section \ref{['sec:sim']}. Right: Distributions of sSFR for mergers and nonmergers.
• Figure 2: Main steps to create a mock F814W CANDELS image from the radiative transferred TNG50 data: 1) The left-most panel shows the image just after it has been processed by SKIRT; 2) Next, we apply an HST F814W filter to the image so we are no longer seeing all wavelengths of light; we also rebin the image to the same pixel scale as the CANDELS mosaics; 3) Next we convolve with the PSF of the telescope to replicate what this galaxy would look like if observed by HST. 4) Lastly we add real CANDELS backgrounds to include real CANDELS noise and background sources. This process will supply a more realistic training set, and is a crucial piece to building a robust network that can classifiy CANDELS galaxies.
• Figure 3: The loss and accuracy curves for our network. The small data set size leads to the bumpier curves, especially in the validation set shown in the orange dashed line. Though the training set curves in solid purple continued to improve, the validation set curves plateaued, so we implemented early stopping to avoid overfitting. We use the weights from epoch 43 as our best model, noted by the grey dashed line.
• Figure 4: The confusion matrix of our Seed 626 network showing that we do classify the majority of galaxies correctly. The darker squares along the diagonal show the galaxies classified correctly.
• Figure 5: The receiver operating characteristic curve for our test set in Seed 626. The better the network performs, the closer the AUC is to 1. The grey dashed line indicates a network that randomly guesses each time, with an accuracy of 0.5.