<p>The Vera C. Rubin Observatory’s Legacy Survey of Space and Time (LSST) is set to revolutionize astronomy by generating an unprecedented petascale dataset. However, the success of its primary science goals, such as precision cosmology with weak lensing, is critically threatened by atmospheric turbulence, which blurs images and can systematically corrupt the faint cosmological signals. The sheer volume of LSST data—approximately 20 terabytes per night—renders classical iterative deconvolution methods computationally infeasible, while generic deep learning approaches often lack the physical guarantees necessary for high-precision science. This creates an urgent need for a fast, robust, and physically-grounded deconvolution framework. To meet this challenge, we introduce a deep learning model that synergistically combines a dual-domain architecture with physics-informed learning. Our U-Net model incorporates a Fast Fourier Transform (FFT) layer at its input, enabling it to directly “see” and correct the frequency-dependent signature of atmospheric blurring. We train the model with a hybrid loss function that enforces both structural realism, via a Point Spread Function (PSF) consistency term, and photometric accuracy, through a flux conservation constraint. Our final model, HF-UNet, produces good reconstructions on realistically degraded images, accurately recovering key galaxy morphological parameters with exceptional fidelity, achieving low Root Mean Square Error (RMSE) values, for instance, 0.05 for ellipticity and 3.37 for half-light radius with the proposed HF-UNet model. Crucially, it exhibits superior robustness when tested on data with mismatched PSF profiles and varying noise levels. This work presents a scientifically reliable deconvolution framework, offering an enabling technology essential for realizing the full scientific potential of the LSST.</p>