CSST-PSFNet: A Point Spread Function Reconstruction Model for the CSST Based on Deep Learning

This paper presents CSST-PSFNet, a deep learning method for high-fidelity point spread function (PSF) reconstruction developed for the Chinese Space Station Survey Telescope (CSST). The model integrates a residual neural network, a lightweight Transformer architecture, and a variational latent representation to address key challenges in CSST imaging, including severe PSF undersampling, inter-band variability, and smooth spatial variation across the focal plane. Trained and validated on high-resolution star-PSF pairs generated by the CSST Main Survey Simulator, CSST-PSFNet achieves improved pixel-level accuracy and more precise recovery of shape parameters relevant to weak lensing compared to widely used PSFEx. On both the standard test dataset and a blurred dataset representing the upper bound of expected on-orbit PSF degradation, the model achieves a size residual precision below 0.005 and an ellipticity residual precision below 0.002. A weak-label adaptation experiment further shows that the model can recover PSFEx-level performance when the true PSF is unknown, demonstrating robustness in controlled degradation scenarios and weak-label adaptation experiments. These results indicate that CSST-PSFNet provides a flexible and extensible framework for future on-orbit PSF calibration in large-scale CSST surveys, with potential applications in weak-lensing cosmology and precision astrophysical measurements.