Towards detecting Primordial non-Gaussianity in the CMB using Spherical Convolutional Neural Networks

TL;DR

This paper introduces spherical CNNs as a promising tool for detecting primordial non-Gaussianity in the CMB, offering a direct analysis method that could complement traditional estimators and scale to future large datasets.

Contribution

It proposes the use of spherical CNNs for analyzing full-sky CMB maps to detect non-Gaussianity, addressing computational limitations of existing methods.

Findings

Spherical CNNs achieve near-optimal error bounds on simulated data.

DeepSphere CNNs closely match Fisher forecast under noisy and masked conditions.

The approach shows potential for scaling to larger future datasets.

Abstract

This paper explores a novel application of spherical convolutional neural networks (CNNs) to detect primordial non-Gaussianity in the cosmic microwave background (CMB), a key probe of inflationary dynamics. While effective, traditional estimators encounter computational challenges, especially when considering summary statistics beyond the bispectrum. We propose spherical CNNs as an alternative, directly analysing full-sky CMB maps to overcome limitations in previous machine learning (ML) approaches that relied on data summaries. By training on simulated CMB maps with varying amplitudes of non-Gaussianity, our spherical CNN models show promising alignment with optimal error bounds of traditional methods, albeit at lower-resolution maps. While we explore several different architectures, results from DeepSphere CNNs most closely match the Fisher forecast for Gaussian test sets under noisy and masked conditions. Our study suggests that spherical CNNs could complement existing methods of non-Gaussianity detection in future datasets, provided additional training data and parameter tuning are applied. We discuss the potential for CNN-based techniques to scale with larger data volumes, paving the way for applications to future CMB data sets.