Speaker
Description
Reconstructing the initial density field of the universe is crucial for improving cosmological parameter constraints. In this work, we employ a U-Net architecture to reconstruct the initial density field from simulated 21-cm and CO line intensity maps from the Epoch of Reionization (EoR). These tracers provide complementary information, with 21-cm maps capturing low-density neutral regions and CO maps tracing high-density star-forming regions. By combining these two maps, we achieve accurate reconstructions across different ionization fractions. Our results demonstrate that the network effectively recovers both large- and small-scale features of the initial density field, achieving a high cross-correlation with the true field ($\geq 0.75$ for $k \leq 1 {\rm Mpc}^{-1}$). To extract cosmological information, we apply Marginal Neural Ratio Estimation (MNRE) to perform simulation-based parameter inference using the reconstructed density field. We find that post-reconstruction information significantly improves constraints on cosmological parameters, reducing uncertainties in $\sigma_8$ and $n_{\rm s}$ by a factor of 2-3. Additionally, we assess the impact of Gaussian random noise on the low-resolution input maps, showing that while the network remains robust in recovering large-scale features, small-scale structures are more affected. Our results highlight the potential of combining multi-tracer intensity mapping with deep learning and neural inference techniques to enhance cosmological constraints from the EoR.
| AI keywords | Simulation-based inference; U-Net; Pattern-recognition |
|---|
