PUREPath-B: A Tessellated Bayesian Framework for CMB Recovery

Vipin Sudevan^1*, Pisin Chen^2,3,4,5

¹Department of Physics, National Sun Yat-sen University, Kaohsiung, Taiwan
²Leung Center for Cosmology and Particle Astrophysics, National Taiwan University, Taipei, Taiwan
³Graduate Institute of Astrophysics, National Taiwan University, Taipei, Taiwan
⁴Department of Physics and Center for Theoretical Sciences, National Taiwan University, Taipei, Taiwan
⁵Kavli Institute for Particle Astrophysics and Cosmology, SLAC National Accelerator Laboratory, Stanford University, California, USA

* Presenter:Vipin Sudevan, email:vipinsudevan1988@gmail.com

We present PUREPath-B, a novel Bayesian deep-learning framework for reconstructing the Cosmic Microwave Background (CMB) B-mode signal from multi-frequency, foreground-contaminated observations. Our network employs a probabilistic, multi-modal U-Net architecture with probabilistic ResNet blocks in the skip connections, enabling hierarchical encoding of multi-scale correlations across frequencies within a variational Bayesian framework. The output is a per-pixel MultivariateNormalDiag distribution which yields both the reconstructed map and uncertainties. The loss combines KL-divergence regularization with a weighted negative log-likelihood, ensuring balanced learning across B-mode realizations with varying tensor-to-scalar ratios (r). Tests on simulated datasets show a structural similarity index (SSIM) above 0.88 and power spectra in excellent agreement with the input CMB signal after bias correction. Incorporating these into cosmological parameter estimation (CPE) provides good recovery of r and lensing amplitude, demonstrating PUREPath-B’s scalable, uncertainty-aware, capability for next-generation CMB polarization experiments.

Keywords: Cosmic Microwave Background, Foreground Minimization, Deep Learning, Variational Inference, Cosmological Parameter Estimation