| dc.description.abstract | No prior machine-learning classifier for bent RAGNs has used both unlabeled data
and purely visually verified labels, despite the crucial role that wide-angle tail (WAT)
and narrow-angle tail (NAT) radio active galactic nuclei (RAGNs) play as tracers of
dense environments in galaxy groups and clusters. We provide the RGC Python module,
which builds a semi-supervised model that combines two recently curated labeled datasets
containing 639 WATs and NATs from a publicly accessible catalog of visually examined
sources with 20,000 unlabeled RAGNs.
The labeled datasets in RGC were preprocessed using PyBDSF, which keeps them for
comparison, and Photutils, which eliminates spurious sources. To create a reliable semi-
supervised binary model, the underlying classifier combines a supervised E2CNN (E(2)-
equivariant Convolutional Neural Network) with the self-supervised learning framework
BYOL (Bootstrap YOur Latent).
The RGC model reaches peak performance with an accuracy of 88.88% and F1-scores of
0.90 for WATs and 0.85 for NATs when trained and assessed on a dataset free of spurious
sources. The model’s attention patterns point to a potential step toward physics-informed
foundation models that can recognize a variety of AGN physical properties, especially
when there is class imbalance. | en_US |
