Using Neural Networks to Differentiate Newly Discovered BL Lacs and FSRQs among the 4FGL Unassociated Sources Employing Gamma-ray, X-ray, UV/Optical and IR Data

TL;DR

This study employs neural networks to classify unassociated gamma-ray sources as FSRQs or BL Lacs using multi-wavelength data, improving the identification process of blazar subclasses in the Fermi catalog.

Contribution

The paper introduces a multi-perceptron neural network classifier that utilizes gamma-ray, X-ray, UV/Optical, and IR data to distinguish between FSRQs and BL Lacs among unassociated sources, enhancing classification accuracy.

Findings

4 FSRQs identified with >99% confidence

50 BL Lacs identified with >99% confidence

58 blazars remain unclassified due to ambiguity

Abstract

Among the ~2157 unassociated sources in the third data release (DR3) of the fourth Fermi catalog, ~1200 were observed with the Neil Gehrels Swift Observatory pointed instruments. These observations yielded 238 high S/N X-ray sources within the 95% Fermi uncertainty regions. Recently, Kerby et al. employed neural networks to find blazar candidates among these 238 X-ray counterparts to the 4FGL unassociated sources and found 112 likely blazar counterpart sources. A complete sample of blazars, along with their sub-classification, is a necessary step to help understand the puzzle of the blazar sequence and for the overall completeness of the gamma-ray emitting blazar class in the Fermi catalog. We employed a multi-perceptron neural network classifier to identify FSRQs and BL Lacs among these 112 blazar candidates using the gamma-ray, X-ray, UV/optical, and IR properties. This classifier provided probability estimates for each source to be associated with one or the other category, such that P_fsrq represents the probability for a source to be associated with the FSRQ subclass. Using this approach, 4 FSRQs and 50 BL Lacs are classified as such with >99% confidence, while the remaining 58 blazars could not be unambiguously classified as either BL Lac or FSRQ.