Machine Learning for Accelerated Cosmological Inference Beyond Flat ΛCDM

Abstract

We review the efficacy of Neural Networks (NN) for emulating cosmological observables (angular acoustic scale θ∗ and binned angular power spectra Db of the CMB temperature anisotropies across ℓ ∈ [30, 400]), and apply our emulator-based inference framework to the nonflat ΛCDM model, in the range ΩK ∈ [−0.2, 0.2]. We achieve speed up actor computation times beyond 104 with respect to standard boltzamn solvers, while maintaining all NN prediction error values below 0.7% relative to current experimental uncertainties, enabling for computationally feasible parameter inference pipelines. We also provide constraints on the curvature density parameter ΩK and other ΛCDM parameters using the NN developed in this work and the temperature maps from Planck 2018 SMICA data. A marginal posterior using our NN model gives curvature density ΩK = 0.001+0.030 −0.070, consistent with a flat Universe.