X-rays: a 3D lattice distinguishes only galaxy clusters and active nuclei, without manual subtraction of the background

On simulated observations of XMM-Newton, a CNN working directly on photon-by-photon event cubes proves to be the most robust classifier for extended and weak X-ray sources.

Differentiating between an extended and faint X-ray source that is a cluster of galaxies and another that is an active galaxy nucleus (AGN) remains a delicate exercise in large surveys. A study submitted on arXiv on October 8 compares several convolutional network architectures trained on simulated observations from the XMM-Newton Space Telescope, with different representations of the same data: spectrum alone (1D), image alone (2D), combination of the two, or cube of space-spectral events (3D).

What the model learned

Better spectral resolution improves results, and the image-+spectrum association already provides a solid foundation. 3D CNN, applied directly to cubes, offers the most stable performance from one cross-validation to the next, although its gain over the 1D+2D combination remains modest. The 3D-GradCAM analysis shows that it learns on its own to separate diffuse intracluster gas thermal emission from localized non-thermal signatures of AGNs, by exploiting correlations lost in projections.

Above all, by working on the total number of photons, the approach implicitly characterizes the background noise, which avoids the subtraction and manual modeling step. The authors see it as an end-to-end chain adapted to the automatic classification of sources in the catalogues of future large X-ray surveys.