Artificial neural networks for selection of pulsar candidates from the radio continuum surveys

Abstract

Pulsar search with time-domain observation is very computationally expensive and data volume will be enormous with the next generation telescopes such as the Square Kilometre Array. We apply artificial neural networks (ANNs), a machine learning method, for efficient selection of pulsar candidates from radio continuum surveys, which are much cheaper than time-domain observation. With observed quantities such as radio fluxes, sky position and compactness as inputs, our ANNs output the "score" that indicates the degree of likeliness of an object to be a pulsar. We demonstrate ANNs based on existing survey data by the TIFR GMRT Sky Survey (TGSS) and the NRAO VLA Sky Survey (NVSS) and test their performance. Precision, which is the ratio of the number of pulsars classified correctly as pulsars to that of any objects classified as pulsars, is about 96$\%$. Finally, we apply the trained ANNs to unidentified radio sources and our fiducial ANN with five inputs (the galactic longitude and latitude, the TGSS and NVSS fluxes and compactness) generates 2,436 pulsar candidates from 456,866 unidentified radio sources. These candidates need to be confirmed if they are truly pulsars by time-domain observations. More information such as polarization will narrow the candidates down further.