Eliminating Primary Beam Effect in Foreground Subtraction of Neutral Hydrogen Intensity Mapping Survey with Deep Learning

In neutral hydrogen (H i ) intensity mapping (IM) survey, foreground contamination on cosmological signal is extremely severe, and systematic effects caused by radio telescopes further aggravate the difficulties in subtracting foreground. We investigate whether the deep-learning method, the 3D U-Net algorithm, can play a crucial role in foreground subtraction when considering the systematic effect caused by the telescope’s primary beam. We consider two beam models, i.e., the Gaussian beam and Cosine beam models. The traditional principal component analysis (PCA) method is employed as a preprocessing step for the U-Net method to reduce the map dynamic range. We find that in the case of the Gaussian beam, the PCA method can effectively clean the foreground. However, the PCA method cannot handle the systematic effect induced by the Cosine beam, and the additional U-Net method can improve the result significantly. To show how well the PCA and U-Net methods can recover the H i signal, we also derive the H i angular power spectrum and H i 2D power spectrum after performing foreground subtraction. It is found that in the case of Gaussian beam, the concordance with the original H i map using U-Net is better than that using PCA by 27.4%, and in the case of Cosine beam, the concordance using U-Net is better than that using PCA by 144.8%. Therefore, the U-Net–based foreground subtraction can efficiently eliminate the telescope primary beam effect and shed new light on recovering H i power spectrum for future H i IM experiments.