RGC: a radio AGN classifier based on deep learning. I. A semi-supervised model for the VLA images of bent radio AGNs

TL;DR

This work introduces RGC, a semi-supervised framework for classifying bent radio AGNs into WAT and NAT morphologies using FIRST/VLA imagery. It combines BYOL self-supervision on a large unlabeled set with E2CNN-based supervised fine-tuning on two curated labeled datasets (R_L1 with spurious sources and R_L2 cleaned of spurious signals), achieving a peak accuracy of 88.88% (R_L2) and strong per-class F1 scores (0.90 for WAT, 0.85 for NAT). The authors release the Sasmal-ML catalog and a Python package, and demonstrate that removing spurious sources improves discriminative ability (ROC-AUC up to 0.94 and PR-AUC up to 0.96 for WATs) and calibration (ECE ≈0.150). They also provide Grad-CAM explanations, revealing that WATs drive attention toward cores and edges while NATs show more complex attention patterns, highlighting both model transparency and astrophysical insights. The results seed a path toward physics-informed foundation models capable of broader RAGN property inference across current and future radio surveys (e.g., EMU, LoTSS, SKA).

Abstract

Wide-angle tail (WAT) and narrow-angle tail (NAT) radio active galactic nuclei (RAGNs) are key tracers of dense environments in galaxy groups and clusters, yet no machine-learning classifier of bent RAGNs has been trained using both unlabeled data and purely visually inspected labels. We release the RGC Python package, which includes two newly preprocessed labeled datasets of 639 WATs and NATs derived from a publicly available catalog of visually inspected sources, along with a semi-supervised RGC model that leverages 20,000 unlabeled RAGNs. The two labeled datasets in RGC were preprocessed using PyBDSF which retains spurious sources, and Photutils which removes them. The RGC model integrates the self-supervised framework BYOL (Bootstrap YOur Latent) with the supervised E2CNN (E2-equivariant Convolutional Neural Network) to form a semi-supervised binary classifier. The RGC model, when trained and evaluated on a dataset devoid of spurious sources, reaches peak performance, attaining an accuracy of 88.88% along with F1-scores of 0.90 for WATs and 0.85 for NATs. The model's attention patterns amid class imbalance suggest that this work can serve as a stepping stone toward developing physics-informed foundation models capable of identifying a broad range of AGN physical properties.

Figures (10)

• Figure 1: An example of the process of masking an image as described in Section \ref{['sec:maskpy']} and \ref{['sec:maskph']}. $\mathbf{R}_n$ is the normalized original image, $\mathbf{M}_{L1}$ its mask created using $\texttt{PyBDSF}$ and $\mathbf{M}_L$ its mask produced through Photutils.
• Figure 2: Left: The architecture of BYOL used in our self-supervised pre-training, adapted from Grill2020. The encoder within BYOL in our case is E2CNN which has been used for our supervised downstream fine-tuning. Right: The architecture of E2CNN as used by Scaife2021, adapted from the LeNet-5 architecture of Lecun1998. The operations GC, MP, GP and FC stand for group convolution, max-pooling, group-pooling and fully-connected, respectively.
• Figure 3: Confusion matrices for the classification of bent RAGNs into WATs and NATs, showing the number of true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) for the RGC model trained on the datasets $\mathbf{R}_{L1}$ (left) and $\mathbf{R}_{L2}$ (right).
• Figure 4: Left: ROC curves and AUC scores for the RGC model trained and tested on two datasets $\mathbf{R}_{L1}$ (red) and $\mathbf{R}_{L2}$ (green), for both WATs (solid) and NATs (dashed), in comparison to the random guess (blue). Right: Precision-Recall (PR) curves and the corresponding AUC-PR scores for the same datasets and sources.
• Figure 5: ECE curves (described in Section \ref{['sec:calibration']}) for our RGC model trained and tested on two datasets for classifying WATs and NATs. The colors and linestyles are the same as Fig. \ref{['fig:roc-pr']}.