Abstract
Deep learning has emerged as a transformative methodology in moderncosmology, providing powerful tools to extract meaningful physical informationfrom complex astronomical datasets. This paper implements a novel Bayesiangraph deep learning framework for estimating key cosmological parameters in aprimordial magnetic field (PMF) cosmology directly from simulated CosmicMicrowave Background (CMB) maps. Our methodology utilizes DeepSphere, aspherical convolutional neural network architecture specifically designed torespect the spherical geometry of CMB data through HEALPix pixelization. Toadvance beyond deterministic point estimates and enable robust uncertaintyquantification, we integrate Bayesian Neural Networks (BNNs) into theframework, capturing aleatoric and epistemic uncertainties that reflect themodel confidence in its predictions. The proposed approach demonstratesexceptional performance, achieving $R^{2}$ scores exceeding 0.89 for themagnetic parameter estimation. We further obtain well-calibrated uncertaintyestimates through post-hoc training techniques including Variance Scaling andGPNormal. This integrated DeepSphere-BNNs framework not only delivers accurateparameter estimation from CMB maps with PMF contributions but also providesreliable uncertainty quantification, providing the necessary tools for robustcosmological inference in the era of precision cosmology.