PUREPath: A Deep Latent Variational Model for Estimating CMB Posterior over Large Angular Scales of the Sky

Abstract

We present a comprehensive neural architecture, the PUREPath, which leverages a nested Probabilistic multi-modal U- Net framework, augmented by the inclusion of probabilistic ResNet blocks in the Expanding Pathway of the decoders, to estimate the posterior density of the Cosmic Microwave Background (CMB) signal conditioned on the observed CMB data and the training dataset. By seamlessly integrating Bayesian statistics and variational methods our model effectively minimizes foreground contamination in the observed CMB maps. The model is trained using foreground and noise contaminated CMB temperature maps simulated at Planck LFI and HFI frequency channels 30 - 353 GHz using publicly available Code for Anisotropies in the Microwave Background (CAMB) and Python Sky Model (PySM) packages. During training, our model transforms initial prior distribution on the model parameters to posterior distributions based on the training data. From the joint full posterior of the model parameters, during inference, a predicitve CMB posterior and summary statistics such as the predictive mean, variance etc of the cleaned CMB map is estimated. The predictive standard deviation map provides a direct and interpretable measure of uncertainty per pixel in the predicted mean CMB map. The cleaned CMB map along with the error estimates can be used for more accurate measurements of cosmological parameters and other cosmological analyses.