Constraining the polarisation flux density and angle of point sources by training a convolutional neural network

¹ Departamento de Física, Universidad de Oviedo, C. Federico García Lorca 18, 33007 Oviedo, Spain
² Instituto Universitario de Ciencias y Tecnologías Espaciales de Asturias (ICTEA), C. Independencia 13, 33004 Oviedo, Spain
e-mail: casasjm@uniovi.es
³ Departamento de Informática, Universidad de Oviedo, Edificio Departamental 1, Campus de Viesques s/n, 33204, Gijón, Spain
⁴ Escuela de Ingeniería de Minas, Energía y Materiales Independencia 13, 33004 Oviedo, Spain

Received: 5 July 2022
Accepted: 24 December 2022

Abstract

Context. Constraining the polarisation properties of extragalactic point sources is a relevant task not only because they are one of the main contaminants for primordial cosmic microwave background B-mode detection if the tensor-to-scalar ratio is lower than r = 0.001, but also for a better understanding of the properties of radio-loud active galactic nuclei.

Aims. We develop and train a machine learning model based on a convolutional neural network to learn how to estimate the polarisation flux density and angle of point sources embedded in cosmic microwave background images knowing only their positions.

Methods. To train the neural network, we used realistic simulations of patches of 32 × 32 pixels in area at the 217 GHz Planck channel with injected point sources at their centres. The patches also contain a realistic background composed of the cosmic microwave background signal, the Galactic thermal dust, and instrumental noise. We split our analysis into three parts: firstly, we studied the comparison between true and estimated polarisation flux densities for P, Q, and U simulations. Secondly, we analysed the comparison between true and estimated polarisation angles. Finally, we studied the performance of our model with the 217 GHz Planck map and compared our results against the detected sources of the Second Planck Catalogue of Compact Sources (PCCS2).

Results. We find that our model can be used to reliably constrain the polarisation flux density of sources above the 80 mJy level. For this limit, we obtain relative errors of lower than 30% in most of the flux density levels. Training the same network with Q and U maps, the reliability limit is above ±250 mJy when determining the polarisation angle of both Q and U sources. Above that cut, the network can constrain angles with a 1σ uncertainty of ±29° and ±32° for Q and U sources, respectively. We test this neural network against real data from the 217 GHz Planck channel, obtaining similar results to the PCCS2 for some sources; although we also find discrepancies in the 300–400mJy flux density range with respect to the Planck catalogue.

Conclusions. Based on these results, our model appears to be a promising tool for estimating the polarisation flux densities and angles of point sources above 80 mJy in any catalogue with very small computational time requirements.