Prépublication · 2026
Map-Level Inference of Primordial Gravitational Waves from the CMB with Spherical Neural Networks
Anto I. Lonappan, Adrian E. Bayer et al. — Monde
Source : arXiv
Résumé
We train a spherical graph convolutional network to infer the tensor-to-scalar ratio $r$ directly from multifrequency cosmic microwave background (CMB) $B$-mode maps on the cut sky. We test the network on simulations of a generic ground-based experiment including lensed CMB, inhomogeneous noise, and stochastic Galactic dust and synchrotron emission. We compare the network with a multifrequency bandpower likelihood built from the same 20,000 simulations used to train it. Working directly on the maps, with no power-spectrum estimation or component separation, the network reduces the root-mean-sq…