Galaxy Mergers in UNIONS -- II: Predicting Timescales in the Post-Merger Regime

TL;DR

The paper tackles the challenge of constraining post-merger timescales by extending the Mummi framework to predict $T_{PM}$ across four bins up to 1.76 Gyr after coalescence using realism-enhanced IllustrisTNG mock images. An ensemble of CNN-ViT models is trained on 256×256 mock images to classify post-mergers into discrete time bins, with a probability-flag mechanism improving precision. Applied to the UNIONS survey, the authors release a catalog of 8,716 post-merger galaxies with $M_*/M_\\odot \ge 10^{10}$ in the range $0.03<z<0.3$, facilitating studies of merger-driven evolution such as star formation, quenching, and AGN activity over the post-merger timeline. The work demonstrates a practical framework for linking morphological post-merger signatures to physical timescales, offering a path toward broader application at higher redshift and in pre-merger phases.

Abstract

Galaxy mergers are critical events that influence galaxy evolution by driving processes such as enhanced star formation, quenching, and active galactic nucleus (AGN) activity. However, constraining the timescales over which these processes occur in the post-merger phase has remained a significant challenge. This study extends the MUlti-Model Merger Identifier (\textsc{Mummi}) framework to predict post-merger timescales ($T_{PM}$) for galaxies, leveraging machine learning models trained on realism-enhanced mock observations derived from the IllustrisTNG simulations. By classifying post-merger galaxies into four temporal bins spanning 0 to 1.76 Gyr after coalescence, \textsc{Mummi} achieves time classification accuracies exceeding 70 per cent. We apply this framework to the Ultraviolet Near Infrared Optical Northern Survey (UNIONS), yielding a catalog of 8,716 post-merger galaxies with $T_{PM}$ predictions and stellar masses $\log(M_*/M_\odot) \geq 10$ at redshifts 0.03 < z < 0.3. These results provide a robust methodology to connect galaxy interaction timescales with physical processes, enabling detailed studies of galaxy evolution in the post-merger regime.

Figures (10)

• Figure 1: IllustrisTNG post-merger sample statistics. We show histograms for the physical properties of our mock sample of UNIONs-like simulated galaxies. Redshift, stellar masses and gas fractions are shown above, while stellar mass ratios, true time since latest merger event and the separation of the closest companion in 3D space in the simulations can be found on the bottom row. Our post-merger sample displays a wide range of properties, from wet to dry mergers, encompassing minor mergers and major mergers.
• Figure 2: Snapshots of simulated post-merger galaxies from IllustrisTNG, highlighting the temporal evolution of morphological features across four post-merger time bins. Each snapshot corresponds to a representative galaxy classified in the respective bin: 0 < $T_{PM}$ < 0.16 Gyr (green outline), 0.16 < $T_{PM}$ < 0.48 Gyr (blue outline), 0.48 < $T_{PM}$ < 0.96 Gyr (yellow outline), and 0.96 < $T_{PM}$ < 1.76 Gyr (red outline). Early post-mergers exhibit prominent morphological disturbances, such as tidal tails and asymmetries, which diminish in intensity with time. By the final bin, galaxies appear largely relaxed, reflecting the gradual fading of merger-induced features.
• Figure 3: Confusion matrix showing the performance of Mummi Step 3 in predicting the number of simulation snapshots that have passed since the merger event. The true labels (vertical axis) and predicted labels (horizontal axis) correspond to the number of snapshots elapsed. Diagonal elements represent the fractions of galaxies correctly classified for a given snapshot count, while off-diagonal elements indicate misclassifications. The matrix highlights Mummi's strong performance for early snapshots (e.g., snapshot 0 with 84% accuracy) and intermediate snapshots, while later stages (e.g., snapshots 8–10) show increasing confusion, reflecting the gradual fading of merger-induced features and the difficulty in distinguishing relaxed systems.
• Figure 4: Confusion matrix showing the performance of Mummi before the application of probability flags in predicting post-merger timescales across four temporal bins: $0 \ \rm Gyr < T_{PM} < 0.16 \ \rm Gyr$, $0.16 \ \rm Gyr < T_{PM} < 0.48 \ \rm Gyr$, $0.48 \ \rm Gyr< T_{PM} < 0.96 \ \rm Gyr$, and $0.96 \ \rm Gyr < T_{PM} < 1.76 \ \rm Gyr$. Each cell represents the fraction of galaxies with a true temporal bin label (vertical axis) assigned to a predicted bin (horizontal axis). Diagonal elements highlight correctly classified galaxies, with classification purity ranging from 47% to 68%. Misclassifications (off-diagonal elements) are more prominent in intermediate bins, reflecting the challenges of distinguishing transitional stages. This matrix provides a baseline for evaluating the improvements introduced by subsequent probability flagging. For ease of comparison with Figs. \ref{['fig:CMFULL']} and \ref{['fig:cm-small']}, all confusion matrices in this work are normalized to show purity in the diagonal elements.
• Figure 5: Accuracy as a function of snapshot confidence intervals for the Mummi framework. Each curve represents a specific temporal bin: Bin 1 $(0 \ \rm Gyr < T_{PM} < 0.16 \ \rm Gyr)$, Bin 2 $(0.16 \ \rm Gyr < T_{PM} < 0.48 \ \rm Gyr)$, Bin 3 $(0.48 \ \rm Gyr < T_{PM} < 0.96 \ \rm Gyr)$, and Bin 4 $(0.96 \ \rm Gyr < T_{PM} < 1.76 \ \rm Gyr)$. Wider bins capture more generalized temporal trends, resulting in increased classification accuracy for all bins. However, the loss of finer temporal resolution is evident, particularly for Bin 2 and Bin 3, which show declining accuracy for narrow bin widths due to overlapping morphological features across adjacent time intervals. This figure highlights the trade-off between temporal precision and classification accuracy in the MUMMI framework, emphasizing the importance of optimizing bin widths for specific scientific objectives.