Fast End-to-End Framework for Cosmological Parameter Inference from CMB Data Using Machine Learning
Larissa Santos, Camila P. Novaes, Elisa G. M. Ferreira, Carlo Baccigalupi · 2025
Read Fast End-to-End Framework for Cosmological Parameter Inference from CMB Data Using Machine Learning by Larissa Santos, Camila P. Novaes, Elisa G. M. Ferreira, Carlo Baccigalupi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-01291.txt.
Introduction
Larissa Santos, Camila Novaes, Elisa Ferreira, and Carlo Baccigalupi build a fast machine learning pipeline for reading cosmology from the cosmic microwave background. The faint polarization signal that tests inflation is buried under much brighter galactic foreground emissions, so cleaning the sky maps is the costly first step of any analysis.
Their end to end framework pairs an analytical blind separation method for foreground removal with a neural network that estimates cosmological parameters directly from full sky model runs. The tests used real model runs for the planned LiteBIRD and PICO satellite missions. The method recovered the optical depth and the tensor to scalar ratio close to their input values, with small error bars. Its LiteBIRD forecasts agree well with the team's own latest numbers.
Catalog record
| CREATOR | Larissa Santos, Camila P. Novaes, Elisa G. M. Ferreira, Carlo Baccigalupi |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 19 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-01291
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Fast End-to-End Framework for Cosmological Parameter Inference from CMB Data Using Machine Learning Preprint: APS/123-QED Larissa Santos, Camila P. Novaes, Elisa G. M. Ferreira, Carlo Baccigalupi Abstract Precise estimation of cosmological parameters from the cosmic microwave background (CMB) remains a central goal of modern cosmology and a key test of inflationary physics. However, this task is fundamentally limited by strong foreground contamination, primarily from Galactic emissions, which obscure the faint CMB B-mode polarization signal. In this Letter, we introduce a fast, simulation-based, end-to-end pipeline that integrates a robust component separation technique with machine-learning, leading to cosmological parameter estimation. Our approach combines the Analytical Blind Separation (ABS) method for foreground removal with a neural network (NN) framework optimized to extract cosmological parameters directly from full-sky simulations. We assess the performance of this methodology for the forthcoming LiteBIRD and PICO satellite missions, designed to detect CMB B-modes with unprecedented sensitivity. Applying the pipeline to realistic sky simulations, we obtain 1 σ 1\sigma errors of 0.0035 (LiteBIRD) and 0.0030 (PICO) for the optical depth τ \tau , and 0.005 (LiteBIRD) and 0.0014 (PICO) for the tensor-to-scalar ratio, r r . In all cases, the recovered parameters are consistent with input values within 1 σ 1\sigma across most of the parameter space. Results for LiteBIRD are in excellent agreement with the latest forecasts from the collaboration. Our findings establish this combined ABS–NN pipeline as a competitive, accurate, and computationally efficient alternative for cosmological parameter inference, offering a powerful framework for forthcoming CMB experiments. I Introduction The cosmic microwave background (CMB), relic radiation from when the Universe was about 380,000 years old, encodes fundamental information about the early Universe. High-precision measurements of its temperature and polarization anisotropies provide stringent tests of cosmological models and the physics driving cosmic inflation. Upcoming CMB space missions, such as LiteBIRD Hazumi et al. 2020 , PICO Hanany et al. 2019 , and CORE Delabrouille 2018 , aim to measure B-mode polarization with unprecedented accuracy, probing the energy scale of inflation. A secondary goal is to achieve cosmic-variance–limited E-mode measurements at low multipoles, improving constraints on the optical depth to reionization, \tau , one of the least precisely determined \Lambda CDM parameters. At microwave frequencies, however, the CMB signal is strongly contaminated by foreground emissions from the Galactic interstellar medium and extragalactic sources. Accurate cosmological inference thus depends on effective foreground modeling and removal. Component separation techniques based on linear combinations of multifrequency data exploit two key properties: (i) distinct frequency dependences of astrophysical emissions and (ii) their statistical independence. Modeling the sky as a mixture of emission components allows component separation through inversion of a linear system. However, residual contamination in maps and power spectra remains a persistent challenge for future experiments. Beyond foreground cleaning, precise cosmological parameter estimation and model testing remain central goals of CMB research. Given current tensions within the \Lambda CDM framework Di Valentino et al. 2025 , alternative inference approaches are both timely and essential. Machine learning (ML), especially neural networks (NNs), has recently emerged as a powerful alternative to traditional likelihood-based methods, enabling simulation-based inference that avoids assumptions about the functional form of the likelihood of the data Cranmer et al. 2020 ; Alsing et al. 2018 ; Alsing and Wandelt 2019 ; Jeffrey et al. 2021 . However, NN-based inference requires extensive training datasets representative of the data distribution, making the repeated application of component separation methods across large ensembles of simulations computationally prohibitive. In this Letter, we propose a fast end-to-end methodology combining cosmological parameter inference via ML with component separation method applied consistently to all simulations. Specifically, we employ the Analytical Blind Separation (ABS) technique ( Yao et al. 2018 ; Zhang et al. 2019 ; Santos et al. 2019 ) to estimate cleaned CMB E- and B-mode power spectra, coupled with a NN architecture optimized to infer cosmological parameters. We demonstrate this framework on simulated LiteBIRD and PICO observations, focusing on two key parameters for next-generation CMB polarization data: the optical depth \tau and the tensor-to-scalar ratio r . II Analysis II.1 Simulations We generate synthetic polarization maps using the Planck Sky Model (PSM) Delabrouille et al. 2013 , which simulates CMB, polarized foregrounds, and instrumental noise (Gaussian noise, uncorrelated from pixel to pixel and from channel to channel) according to LiteBIRD and PICO specifications (Tables 1 and 2 ). All maps are produced at a HEALPix resolution of N_{side}=256 Górski et al. 2005 . II.1.1 CMB maps We compute theoretical EE and BB angular power spectra using the Boltzmann solver CLASS Lesgourgues 2011 ; Blas et al. 2011 under the \Lambda CDM framework. To train and test the NN for cosmological inference, we calculate C_{\ell}^{EE} , C_{\ell}^{BB} for a set of 700 different { r , \tau } pairs of values (each pair refereed as a cosmology). The training set spans 0<r<0.05 and 0.01<\tau<0.13 (500 samples), while the test set covers 0<r<0.03 and 0.038<\tau<0.070 (200 samples). This strategy ensures that the NNs are trained over sufficiently broad parameter ranges, mitigating the prior edge effects that could otherwise introduce biases in the inferred parameters Villaescusa-Navarro et al. 2022 ; Perez et al. 2022 . The narrower parameter range adopted for the test set, while avoiding such boundary-induced artifacts, confines the inference to physically plausible regions of the parameter space consistent with current cosmological constraints Aghanim et al. 2020 . We sample the parameter space using the Latin Hypercube approach McKay et al. 1979 and generate 10 realizations of Q and U polarization maps for each cosmology using the power spectra. All other \Lambda CDM parameters are fixed to the Planck 2018 best-fit values: H_{0}=67.27 km s -1 , \Omega_{c}=0.363 , \Omega_{b}=0.049 , n_{s}=0.9649 . Given the constraint 10^{9}A_{s}e^{-2\tau}=1.884 Wolz et al. 2023 , we fix this combination while varying A_{s} and \tau . The resulting E - and B -mode spectra (Figure 1 , left) illustrate the impact of varying r and \tau . Figure 1: Angular power spectra of the input clean CMB simulations (left) and the spectra recovered by ABS from LiteBIRD- (middle) and PICO-like (right) simulations. ABS spectra include the effect of the corresponding largest beam (first row of Tables 1 and 2 ). The green and blue regions represent the noise levels (frequency dependent) for LiteBIRD and PICO cases, respectively. II.1.2 Foreground contaminants In addition to the CMB, several astrophysical processes contribute to sky emission across 21–462 GHz. We model the two dominant sources of polarized Galactic foregrounds: synchrotron radiation from relativistic charged particles in Galactic magnetic fields, dominant at low frequencies, and thermal dust emission from heated dust grains, prevailing at high frequencies. Extragalactic contributions from radio and far-infrared galaxies, along with cluster emission, are also included to ensure a comprehensive description of polarized contaminants. Table 1: Litebird specifications Hazumi et al. 2020 ; Allys et al. 2023 . Band center Beam FWHM noise level (GHz) (arcmin) ( \mu K CMB -arcmin) 040 70.5 37.42 050 58.5 33.46 060 51.1 21.31 068 47.1 16.87 078 43.8 12.07 089 41.5 11.3 100 37.8 6.56 119 33.6 4.58 140 30.8 4.79 166 28.9 5.57 195 28.6 5.85 235 24.7 10.79 280 22.5 13.80 337 20.9 21.95 402 17.9 47.45 Table 2: PICO specifications Hanany