An emulator-based forecasting on astrophysics and cosmology with 21 cm and density cross-correlations during EoR

Abstract

The 21 cm signal arising from fluctuations in the neutral hydrogen field, and its cross-correlation with other tracers of cosmic density, are promising probes of the high-redshift Universe. In this study, we assess the potential of the 21 cm power spectrum, along with its cross power spectrum with dark matter density and associated bias, to constrain both astrophysics during the reionization era and the underlying cosmology. Our methodology involves emulating these estimators using an Artificial Neural Network (ANN), enabling efficient exploration of the parameter space. Utilizing a photon-conserving semi-numerical reionization model, we construct emulators at a fixed redshift ($z = 7.0$) for $k$-modes relevant to upcoming telescopes such as SKA-Low. We generate $\sim7000$ training samples by varying both cosmological and astrophysical parameters along with initial conditions, achieving high accuracy when compared to true simulation outputs. While forecasting, the model involves five free parameters: three cosmological ($Ω_m$, $h$, $σ_8$) and two astrophysical (ionizing efficiency, $ζ$, and minimum halo mass, $M_{\mathrm{min}}$). Using a fiducial model at the mid-reionization stage, we create a mock dataset and perform forecasting with the trained emulators. Assuming a 5% observational uncertainty combined with emulator error, we find that the 21 cm and 21 cm-density cross power spectra can constrain the Hubble parameter ($h$) to better than 6% at a confidence interval of 95%, with tight constraints on the global neutral fraction ($Q_{\mathrm{HI}}$). The inclusion of bias information further improves constraints on $σ_8$ (< 10% at 95% confidence). Finally, robustness tests with two alternate ionization states and a variant with higher observational uncertainty show that the ionization fractions are still reliably recovered, even when cosmological constraints weaken.

Figures (9)

• Figure 1: Comparison of true 21 cm power spectrum and corresponding predicted estimates using ANN at different $k$ bins used in this work. The black points correspond to test dataset while the red line signifies True=Prediction. The corresponding $R^2$ value is 0.98.
• Figure 2: Comparison of true cross power amplitude between 21 cm and $\delta_m$ field with the corresponding predicted estimates using ANN at different $k$ bins used in this work. Other descriptions are similar to Figure \ref{['fig:comp_21_pow']}. This corresponds to an $R^2$ value of 0.99.
• Figure 3: Comparison of true 21 cm bias and the corresponding predicted estimates using ANN at different $k$ bins used in this work. Other descriptions are similar to Figure \ref{['fig:comp_21_pow']}. The corresponding $R^2$ value is 0.92.
• Figure 4: Plots of 21 cm power spectra, 21 cm-density cross power, and its bias for a few random models from the test set. The solid lines are the true models, while the dashed lines are the corresponding predictions.
• Figure 5: Top panel: Snapshots of Density ($\Delta$), collapsed fraction ($f_{\mathrm{coll}}$), and neutral fraction field ($x_{\mathrm{HI}}$), gradually from left to right, for the fiducial model utilized to generate the mock dataset as described in section \ref{['sec:gen_mock']}. Bottom panel: The 21 cm power spectra ($\langle\delta T_b\rangle^2\Delta_{21}^2$), 21 cm and density cross power spectra ($\vert \Delta_{21\times\delta}^2\vert$), and the bias of cross spectra ($b^2_{21\times\delta}$) for the corresponding model, gradually from left to right.