Imaging physics-driven artificial intelligence makes ground-based telescope resolve deep field universe

Astronomical observations are constrained by a long-standing trade-off: ground-based telescopes offer wide-field access at a low cost but suffer from atmospheric distortions, whereas space telescopes provide diffraction-limited clarity at the expense of narrow coverage and high cost. This paper presents StelLens, a physics-driven artificial intelligence model that brings ground-telescope imaging data to the level of space-telescope imaging quality by jointly learning the ground imaging prior, the space imaging prior, and the highly non-linear relationship between ground and space images. Applied to Sloan Digital Sky Survey (SDSS) observations and benchmarked with Hubble Space Telescope (HST) optical images, StelLens achieves a 13× improvement in the median image-quality metric, reduces spurious detections by ~50%, enables 295% more detectable sources, and improves the limiting magnitude by 4.37 magnitude. Extensive experiments also demonstrate accurate reconstruction of geometric properties, such as axis ratios and characteristic sizes. Leveraging this approach, the SDSS archive effectively becomes a wide-field survey with space-telescope-like quality, covering ~14,000 square degrees, nearly 300× the area accessible to HST, without additional cost on the telescope. Our StelLens establishes a new paradigm for astronomical surveys, demonstrating that AI can bridge the gap between ground and space observations, enabling clearer and deeper exploration of the universe at scale.