Deep Learning for Point Spread Function Modeling in Cosmology

TL;DR

This work introduces a data-driven framework for Point Spread Function modeling in cosmology that outperforms the established PIFF approach by learning a cross-FoV, latent PSF representation with an autoencoder and interpolating it across the focal plane using Gaussian Processes. Trained on a large Subaru/HSC dataset of $2.79\times 10^6$ stars, the model encodes $25\times 25$ pixel star images into a 16-d latent space and reconstructs PSFs with a mean squared error of $3.4\times 10^{-6}$, beating PIFF at $3.7\times 10^{-6}$. The latent space shows smooth, physically meaningful variations across the focal plane, and GP interpolation enables continuous PSF estimation across the full FoV. The approach lays groundwork for integration into the LSST Science Pipelines, offering a scalable, data-driven alternative to current PSF modeling for precise weak lensing measurements and cosmological inference.

Abstract

We present the development of a data-driven, AI-based model of the Point Spread Function (PSF) that achieves higher accuracy than the current state-of-the-art approach, "PSF in the Full Field-of-View'' (PIFF). PIFF is widely used in leading weak-lensing surveys, including the Dark Energy Survey (DES), the Hyper Suprime-Cam (HSC) Survey, and the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST). The PSF characterizes how a point source, such as a star, is imaged after its light traverses the atmosphere and telescope optics, effectively representing the "blurred fingerprint'' of the entire imaging system. Accurate PSF modeling is essential for weak gravitational lensing analyses, as biases in its estimation propagate directly into cosmic shear measurements -- one of the primary cosmological probes of the expansion history of the Universe and the growth of large-scale structure for dark energy studies. To address the limitations of PIFF, which constructs PSF models independently for each CCD and therefore loses spatial coherence across the focal plane, we introduce a deep-learning-based framework for PSF reconstruction. In this approach, an autoencoder is trained on stellar images obtained with the Hyper Suprime-Cam (HSC) of the Subaru Telescope and combined with a Gaussian process to interpolate the PSF across the telescope's full field of view. This hybrid model captures systematic variations across the focal plane and achieves a reconstruction error of $3.4 \times 10^{-6}$ compared to PIFF's $3.7 \times 10^{-6}$, laying the foundation for integration into the LSST Science Pipelines.

Figures (6)

• Figure 1: Description of the focal plane in the telescope.
• Figure 2: Architecture of the autoencoder.
• Figure 3: Comparison between PSF reconstructions using the traditional PIFF model and the Autoencoder-based method (sample from the full dataset).
• Figure 4: Corner plot of the 16 latent dimensions of the autoencoder. The diagonal panels show the marginal distributions of each latent variable, while the off-diagonal panels display the pairwise correlations.
• Figure 5: Spatial distribution of the field of view (FoV) projected onto each of the latent dimensions of the autoencoder. Each panel shows how the corresponding latent variable captures variations of the PSF across the focal plane.