Estimation of the Full-sky Power Spectrum between Intermediate and Large Angular Scales from Partial-sky CMB Anisotropies Using an Artificial Neural Network

Abstract

Reliable extraction of cosmological information from observed cosmic microwave background (CMB) maps may require removal of strongly foreground-contaminated regions from the analysis. In this paper, we employ an artificial neural network (ANN) to predict the full-sky CMB angular power spectrum between intermediate and large angular scales from the partial-sky spectrum obtained from a masked CMB temperature anisotropy map. We use a simple ANN architecture with one hidden layer containing 895 neurons. Using 1.2 × 105 training samples of full-sky and corresponding partial-sky CMB angular power spectra at HEALPix pixel resolution parameter N side = 256, we show that the spectrum predicted by our ANN agrees well with the target spectrum at each realization for the multipole range 2 ≤ l ≤ 512. The predicted spectra are statistically unbiased, and they preserve the cosmic variance accurately. Statistically, the differences between the mean predicted and underlying theoretical spectra are within approximately 3σ. Moreover, the probability densities obtained from predicted angular power spectra agree very well with those obtained from “actual” full-sky CMB angular power spectra for each multipole. Interestingly, our work shows that the significant correlations in input cut-sky spectra due to mode–mode coupling introduced on the partial sky are effectively removed, since the ANN learns the hidden pattern between the partial- and full-sky spectra preserving all of the statistical properties. The excellent agreement of statistical properties between the predicted and the ground truth demonstrates the importance of using artificial intelligence systems in cosmological analysis more widely.