Blazar classification from multi-wavelength data using Deep Learning

TL;DR

The paper tackles blazar subclassification by training a deep feed-forward neural network to distinguish BL Lac objects from FSRQs using multi-wavelength data sourced from 4FGL-DR4, ROMA-BZCAT, and 3LAC. It compares two datasets: a four-feature setup and an expanded seven-feature setup, evaluated with five-fold cross-validation and 700 epochs per fold, achieving high accuracy (≈0.92 and ≈0.91, respectively) and strong ROC/PR performance with an AUC around 0.977. The study demonstrates that even a comparatively simple architecture can leverage cross-band information to classify BCUs, while highlighting the dependence on complete multi-wavelength coverage and proposing future interpretability analyses. The workflow and results support scalable, data-driven blazar population studies and can inform follow-up observations and CTA-era surveys by refining blazar catalogs through multi-wavelength machine learning.

Abstract

The launch of the Fermi-LAT telescope has revolutionized gamma-ray astronomy by detecting over 7,000 gamma-ray emitting objects. A major fraction of the objects are blazars of known type, and a similar fraction of objects were classified as blazars of uncertain types (BCUs). Apart from that, some of the objects were found to be unassociated with any other classes, and no information is available in other wavebands. These types of objects are classified as unassociated objects in the Fermi catalog. To classify the unassociated objects into known categories and BCUs into a known type of blazar, numerous efforts have been made using the machine learning approach. The ideal way of classification would be to have multi-wavelength temporal and spectral information, which is nearly impossible to have for this number of objects in the near future. In this paper, we focus on classifying BCUs into other types of blazars, such as FSRQs and BL Lacs. For this purpose, we have developed for the first time a deep feed-forward Artificial Neural Network (ANN) to classify them, using multi-wavelength data. The complete understanding of blazars can only be known through multi-wavelength observation, and hence, we begin with four input parameters that cover broadband information (radio, optical, X-ray fluxes, and redshift) to train the neural network, and then extend the framework by including additional parameters to examine their impact on the outcome.

Figures (5)

• Figure 1: Feedforward Artificial Neural Network (ANN) architecture used for binary classification of blazar sources. The input layer accepts feature vectors of dimension 4 or 7, depending on the dataset. The network consists of three fully connected layers: the first maps the input to a 16-dimensional space, followed by a ReLU activation; the second compresses this to 8 dimensions with another ReLU activation; and the final layer maps to a 2-dimensional output representing the target classes (BL Lac and FSRQ). The model is trained using the AdamW optimizer with weight decay, a StepLR scheduler, and Cross-Entropy Loss.
• Figure 2: Comparison of confusion matrices showing classification performance for the two datasets. Subfigure (a) shows the confusion matrix plot for Dataset 1, and Subfigure (b) presents the confusion matrix plot for Dataset 2.
• Figure 3: Receiver Operating Characteristic (ROC) curves illustrating the performance of the classification model on two independent datasets. Subfigure (a) shows the ROC curve for Dataset 1, and Subfigure (b) presents the ROC curve for Dataset 2, allowing direct comparison of model effectiveness across different datasets.
• Figure 4: Precision-recall (PR) curves illustrating the performance of the classification model on two independent datasets. Subfigure (a) shows the PR curve for Dataset 1, and Subfigure (b) presents the (PR) curve for Dataset 2, allowing direct comparison of model effectiveness across different datasets.
• Figure 5: Loss curve comparison indicating the convergence behaviour for the two datasets.