Deep learning for Sunyaev–Zel’dovich detection in Planck

V Bonjean

doi:10.1051/0004-6361/201936919

Abstract

The Planck collaboration has extensively used the six Planck HFI frequency maps to detect the Sunyaev–Zel’dovich (SZ) effect with dedicated methods, for example by applying (i) component separation to construct a full-sky map of the y parameter or (ii) matched multi-filters to detect galaxy clusters via their hot gas. Although powerful, these methods may still introduce biases in the detection of the sources or in the reconstruction of the SZ signal due to prior knowledge (e.g. the use of the generalised Navarro, Frenk, and White profile model as a proxy for the shape of galaxy clusters, which is accurate on average but not for individual clusters). In this study, we use deep learning algorithms, more specifically, a U-net architecture network, to detect the SZ signal from the Planck HFI frequency maps. The U-net shows very good performance, recovering the Planck clusters in a test area. In the full sky, Planck clusters are also recovered, together with more than 18 000 other potential SZ sources for which we have statistical indications of galaxy cluster signatures, by stacking at their positions several full-sky maps at different wavelengths (i.e. the cosmic microwave background lensing map from Planck, maps of galaxy over-densities, and the ROSAT X-ray map). The diffuse SZ emission is also recovered around known large-scale structures such as Shapley, A399–A401, Coma, and Leo. Results shown in this proof-of-concept study are promising for potential future detection of galaxy clusters with low SZ pressure with this kind of approach, and more generally, for potential identification and characterisation of large-scale structures of the Universe via their hot gas.

Highlights

In the past decades, new statistical developments have begun to play an important role in data reduction and data analysis
The Planck maps were masked from the Planck Catalogue of Compact Sources (PCCS, Planck Collaboration XXVI 2016), and the ROSAT map was masked from the point sources detected in ROSAT, Chandra, and XMM-Newton (Boller et al 2016; Evans et al 2010; Rosen et al 2016, respectively)
These results show that the U-net trained on SZ sources with a high signal-to-noise ratio (i.e. Planck clusters) has learned the frequency dependency and the spatial features of the SZ effect in the six Planck HFI frequency maps, and that some of the sources detected with the U-net that do not belong to the Planck or the MCXC catalogue might be actual galaxy clusters, even at the lowest bin of associated prediction index p

Summary

Introduction

New statistical developments have begun to play an important role in data reduction and data analysis. Algorithms are designed to work on unlabelled data. The user must in this case have a perfect knowledge of the labels or of the properties of reference that is used as output in the training catalogue. This type includes algorithms such as artificial neural networks (ANN, White & Rosenblatt 1963), random forests (RF, Ho 1995), and support vector machine (SVM, Hearst 1998). With very complex architectures of superposed layers, may enter the category of deep-learning (DL) algorithms, as this is the case for ANN and convolutional neural networks (CNN, Fukushima 1980)

Methods

Results

Conclusion