Deep-learning mitigation of foregrounds and beam effects in 21-cm intensity mapping using hybrid frequency differencing and PCA

TL;DR

This work assesses foreground mitigation for 21-cm intensity mapping under realistic beam effects by comparing FD-based and PCA-based UNet reconstructions and introducing a hybrid dual-channel input. Using CRIME-based simulations with 192 patches and MeerKAT-like Gaussian and Cosine beams, the study shows that FD+UNet and PCA+UNet perform similarly without beam or with Gaussian beam, but both under Cosine beam bias large-scale cross-correlation by ~5–8% at $k<0.1\,h\mathrm{Mpc}^{-1}$. The proposed hybrid dual-channel approach substantially improves recovery, keeping the large-scale cross-power near unity and delivering 5–8% gains over individual channels, particularly under realistic beam distortions. These results underpin robust HI reconstruction for precision BAO and RSD measurements in upcoming low-redshift 21-cm surveys, with future work planned to include noise, calibration, and real data in an end-to-end pipeline.

Abstract

21-cm intensity mapping (IM) is a powerful technique to probe the large-scale distribution of neutral hydrogen (HI) and extract cosmological information such as the baryon acoustic oscillation feature. A key challenge lies in recovering the faint HI signal from bright foregrounds and frequency-dependent beam effects, which can compromise traditional cleaning methods like principal component analysis (PCA) by removing part of the cosmological signal. Deep-learning approaches have recently been proposed to mitigate these effects by learning mappings between contaminated and true cosmological signals. Building upon our previous work~\citep{2024PhRvD.109f3509S} on the frequency-differencing (FD) method, this study extends the framework to systematically compare FD-based and PCA-based UNet reconstructions using realistic simulations that include foregrounds and beam convolution. We find that both approaches perform comparably without beam or with a Gaussian beam, but under a realistic cosine beam they systematically underestimate the large-scale cross-correlation power spectrum, particularly for $k<0.1 h~\mathrm{Mpc}^{-1}$. To address this limitation, we explore a hybrid approach in which the UNet is trained with two input channels, one constructed from FD and the other from PCA cleaning, allowing the network to simultaneously exploit the strengths of both inputs. This two-channel strategy achieves superior performance, maintaining the cross-correlation power spectrum close to unity on large scales under a cosine beam, improving by 5-8% relative to either FD-based or PCA-based UNet alone. These results demonstrate that providing complementary FD and PCA information to a single deep network is an effective route to robust HI reconstruction, laying the groundwork for precision BAO measurements with future low-redshift 21 cm IM surveys.

Figures (17)

• Figure 1: Comparison of beam convolution effects on a simulated point-source map. Left: the input map containing point-like sources. Middle: the convolved map using a Gaussian beam model. Right: the convolved map using the Cosine beam model.
• Figure 2: (Left) Normalized angular beam profiles of the Gaussian (blue) and Cosine (red) beam models as a function of angular separation from the beam center at 1100 MHz. (Right) Frequency dependence of the FWHM for the Gaussian (blue) and Cosine (red) beam models.
• Figure 3: Visualization of the UNet architecture. The input is a cube of size $64^3$ with two channels derived from the frequency-difference and PCA-cleaned maps. The architecture consists of 13 layers, with 6 encoder layers, 6 decoder layers, and 1 bottleneck layer. Each layer includes three convolution operations with batch normalization and activation functions (except for the output layer). The annotations below each layer (e.g., $32 \times 64^2 \times 2$) denotes the number of feature maps, the spatial dimension, and the two-channel processing. Down, right, and up arrows represent max-pooling, skip connections, and transpose convolutions, respectively.This visualization was created using the PlotNeuralNet library.
• Figure 4: Comparison of preprocessing methods applied to pure HI signal fields. From left to right, the panels show the target truth, the PCA-3 reconstruction, and the frequency differencing result, respectively.
• Figure 5: Comparison of preprocessing methods applied to the pure HI fields. Left: cross-correlation coefficient $R_{\text{cross}}(k)$ between the reconstructed and true HI signals. Right: pixel-to-pixel comparison of the preprocessed temperature $T_{\mathrm{diff}}$ with the true temperature $T_{\mathrm{true}}$, where the contours enclose 65% and 95% of the grid cells.