Comparison and verification methods to trace interaction-driven disturbances in galaxies

Technical Summary: Comparison and Verification Methods to Trace Interaction-Driven Disturbances in Galaxies

Problem Statement
The hierarchical formation of cosmological structures under the $\Lambda$CDM model implies that galaxy mergers are frequent, producing low-surface-brightness (LSB) tidal debris (tails, streams, shells) and broader morphological disturbances (asymmetric haloes, double nuclei). While visual classification has historically been the standard for identifying these features, the impending data volume from the Vera C. Rubin Observatory's Legacy Survey of Space and Time (LSST) and the Euclid mission renders manual inspection infeasible. There is an urgent need for robust automated tools to detect these interaction-driven signatures. However, existing automated methods face trade-offs: supervised machine learning (ML) often requires large, high-quality labelled datasets that are scarce for faint tidal features, while traditional non-parametric methods like the Concentration-Asymmetry-Smoothness (CAS) parameters may lack sensitivity to diffuse structures. This paper evaluates two automated approaches—a Self-Supervised Learning (SSL) model and the CAS parameter method—against visual classification to determine their efficacy in tracing merger signatures.

Methodology
The study utilizes a dataset of 1,826 mock images generated from the IllustrisTNG-100 cosmological simulation. To align with the expected depth and resolution of LSST data, the authors re-generated mock images at a median redshift of $z \sim 0.15$ (spanning $0.02 < z < 0.29$), resulting in a resolution of $0.43''$ pixel$^{-1}$ and surface brightness limits comparable to a 10-year LSST survey ($\mu_g \sim 30.3$ mag arcsec$^{-2}$).

The evaluation framework relies on three distinct classification methods:

1. Visual Classification (Benchmark): A team visually classified the re-made mock images into four categories: Streams/Tails, Shells, Asymmetric Stellar Haloes, and Double Nuclei. A confidence system (0–3) was applied, where confidence levels $\ge 2$ were treated as the ground truth for "disturbed" systems.
2. Self-Supervised Learning (SSL): The authors evaluated the SSL model developed by Desmons et al. (2024). This architecture employs a ResNet-20 encoder trained via Nearest Neighbour Contrastive Learning of visual Representations (NNCLR) on unlabelled data to extract low-dimensional features. Crucially, the authors did not retrain the encoder; instead, they fine-tuned only the linear classifier head using a small subset of the labelled mock images (training on 1,440 images, with validation and test splits). The model was further tested on a reduced training set (480 images) and applied to a completely unseen prediction set (960 images).
3. CAS Parameters: The traditional Concentration-Asymmetry-Smoothness method was applied to the 5-band combined mock images. The study focused on the Asymmetry ($A$) and Smoothness ($S$) parameters, applying the standard merger selection criteria of $A > 0.35$ and $A > S$.

Performance was quantified using True Positive Rate (Recall/Completeness), False Positive Rate (Contamination), Precision, F1 Score, and Area Under the ROC Curve (AUC), with visual classification serving as the reference standard.

Key Results

• Visual Classification Baseline: In the re-made mock images (simulating LSST conditions), 25.1% of galaxies exhibited interaction-driven disturbances with high confidence ($\ge 2$). This is lower than the 34.2% found in the original, closer mock images, highlighting the impact of distance and resolution on the detectability of faint LSB structures. Asymmetric haloes were the most common signature, while shells were nearly absent due to surface brightness dimming.
• SSL Model Performance: By retraining only the linear classifier on a small labelled dataset, the SSL model achieved a high ROC AUC of $0.92 \pm 0.01$. At a fixed False Positive Rate (FPR) of 0.2, the model demonstrated high recall (completeness) of $0.86 \pm 0.04$ with a precision of 0.56. The model maintained robust performance even when the training set was reduced to 480 unique galaxies, indicating that the pre-trained encoder effectively generalizes to new data with minimal labelled supervision.
• CAS Method Performance: Using the traditional threshold of $A > 0.35$, the CAS method exhibited high precision (0.77) but very low recall (0.20). It successfully identified a "clean" sample of disturbed systems but failed to recover the majority of galaxies visually classified as disturbed. The method showed a weak correlation with stellar mass compared to the other methods.
• Threshold Optimization: When the CAS asymmetry threshold was optimized via ROC analysis to match the SSL evaluation criteria (FPR = 0.2), the threshold lowered to $A > 0.14$. This improved recall to 0.59 but reduced precision to 0.48. Despite this adjustment, the CAS method still underperformed the SSL model across all metrics (AUC 0.76 vs. 0.90).
• Stellar Mass Correlation: Both visual classification and the SSL model revealed a strong positive correlation between stellar mass and the likelihood of identifying disturbed systems. In contrast, the CAS method showed a negligible correlation, suggesting it is less sensitive to the mass-dependent trends in merger signatures detectable in deep imaging.

Significance and Conclusions
The paper concludes that the SSL approach, specifically the strategy of retraining only the linear classifier on a small labelled dataset while leveraging a pre-trained self-supervised encoder, offers a highly effective solution for automated disturbance detection in large-scale surveys. It provides a superior balance of completeness and precision compared to the traditional CAS method.

While the CAS method remains useful as a conservative tool for identifying high-confidence, clean samples of mergers, it significantly underestimates the fraction of disturbed galaxies, particularly those with faint, diffuse tidal features. The SSL model's ability to achieve high recall ($\sim 86\%$) with low contamination makes it better suited for constructing a complete census of merger-related features in upcoming LSST data. The study emphasizes that the choice of detection method significantly influences the observed relationship between galaxy properties (such as stellar mass) and merger activity, with traditional parametric methods potentially missing the broader population of interacting systems.