Ponente
Descripción
Traditional methods for component separation, while effective, often require extensive masking to mitigate contamination from galactic and extragalactic foregrounds, limiting the amount of usable sky for analysis and introducing residual biases and systematic uncertainties. By employing modern neural network (NN) architectures, such as convolutional neural networks (CNNs) and fully convolutional neural networks (FCNNs), this project aims to overcome these limitations, delivering a more accurate and efficient separation of the CMB signal from foreground contaminants with minimal masking or developing methods capable of directly estimating power spectra. This novel approach not only increases the efficiency of data analysis but also enhances the accuracy of polarization measurements, particularly for the challenging BB-mode signal, which is a key observable for probing inflationary gravitational waves.
Nevertheless, since most FCNNs are designed for 2D image processing, their use for our purposes introduces challenges in capturing large-scale CMB fluctuations, transitioning from patch-based analyses to full-sky representations, and introducing potential polarization leakage effects at patch boundaries. Having this into account, the first part of the research proposed, which is the subject of the present talk, has been devoted to investigate these methodological issues, seeking to optimize patch size and resolution and evaluate reprojection techniques that may mitigate boundary artifacts and improve the accuracy of power spectrum estimation.
