Reconstruction of Reionization Histories from 21 cm Power-Spectrum Evolution with Artificial Neural Networks

Abstract

We investigate whether the redshift evolution of the fixed-$k$ dimensionless 21 cm power spectrum, $\Delta^2_{21}(k, z)$, contains sufficient information to reconstruct reionization histories $x_{\rm HI}(z)$ with artificial neural networks. Using semi-numerical realizations generated within a restricted three-parameter 21cmFAST model family, we train a compact feed-forward network to learn the inverse mapping from power-spectrum trajectories to the neutral-fraction history on a 97-point redshift grid spanning $6 \le z \le 15$. For $k = 0.1$, $0.5$, and $1.0\ h\ \mathrm{Mpc}^{-1}$, representative tests on an independent test set show that the midpoint redshift $z_{50}$ is recovered more accurately than the duration $\Delta z = z_{75} - z_{25}$ in terms of dimensionless relative errors: $z_{50}$ is reconstructed with $\mathrm{MAE} = 0.0046$ and $\mathrm{RMSE} = 0.0100$, whereas $\Delta z$ yields $\mathrm{MAE} = 0.0302$ and $\mathrm{RMSE} = 0.0378$. This result indicates that fixed-$k$ power-spectrum evolution carries stronger information about the timing of reionization than about the detailed width of the transition within the adopted prior. We further test an idealized foreground-free SKA1-Low-like thermal-plus-sample-variance noise model and find that the reconstruction remains stable in the favorable signal-to-noise regime considered here. These results demonstrate that neural networks can serve as prior-dependent inverse mapping for reconstructing reionization histories from 21 cm power-spectrum evolution.