Optimizing neural network techniques in classifying Fermi-LAT gamma-ray sources

TL;DR

This paper presents an optimized neural network approach that significantly improves the classification accuracy of gamma-ray sources detected by Fermi-LAT, reducing unclassified sources and enhancing astrophysical source identification.

Contribution

The study introduces an enhanced neural network model incorporating gamma-ray spectra and multiwavelength data, achieving an 80% improvement in classification performance over previous methods.

Findings

Reduced unclassified blazars from 77 to 15

Classified blazars into BL Lacs and FSRQ in a 2:1 ratio

Achieved 90% precision in classification

Abstract

Machine learning is an automatic technique that is revolutionizing scientific research, with innovative applications and wide use in astrophysics. The aim of this study was to developed an optimized version of an Artificial Neural Network machine learning method for classifying blazar candidates of uncertain type detected by the Fermi Large Area Telescope (LAT) gamma-ray instrument. The initial study used information from gamma-ray light curves present in the LAT 4-year Source Catalog. In this study we used additionally gamma-ray spectra and multiwavelength data, and certain statistical methods in order to improve classification. The final result of this study increased the classification performance by about 80 per cent with respect to previous method, leaving only 15 unclassified blazars instead of 77 out of total 573 in the LAT catalog. Other blazars were classified into BL Lacs and FSRQ in ratio of about two to one, similar to previous study. In both studies a precision value of 90 per cent was used as a threshold for classification.