Neural blind deconvolution to reconstruct high-resolution ground-based solar observations

¹ Institute of Physics, University of Graz, Universitätsplatz 5, 8010 Graz, Austria
² High Altitude Observatory, NSF National Center for Atmospheric Research, 3080 Center Green Dr, Boulder, USA
³ Kanzelhöhe Observatory for Solar and Environmental Research, University of Graz, Treffen am Ossiacher See, Austria
⁴ National Solar Observatory, 3665 Discovery Drive, Boulder, CO 80303, USA

^★ Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Received: 12 January 2026
Accepted: 3 March 2026

Abstract

Context. Ground-based solar observations enable unprecedented spatial, spectral, and temporal resolution of the lower solar atmosphere, yet Earth’s turbulent atmosphere imposes significant limitations, requiring advanced post facto image reconstruction. State-of-the-art reconstruction methods are based on restoring a burst of short exposure frames to a single observation. Limitations of these techniques arise due to the sparse information available concerning the atmospheric point spread function (PSF), which degrades the observations and, consequently, the quality of reconstructions. Developing new reconstruction methods is essential for providing high-quality data products for the study of the lower solar atmosphere on the smallest scales.

Aims. We aim to develop a novel image-reconstruction method to achieve unprecedented spatial resolution from short exposure image bursts. This can provide high-quality reconstructions and therefore advance the study of the smallest spatial scales from the solar photosphere to the chromosphere.

Methods. In this paper, we present a novel approach for high-resolution solar-image reconstruction based on physics-informed neural networks. In the training process, the neural network maps coordinate points (x, y) directly to their corresponding intensity values o(x, y), while simultaneously updating the PSF parameters. The method convolves the "true" object from the neural network with the estimated PSFs and optimizes the network by minimizing the loss between the synthesized and real short-exposure image burst. This approach enabled the simultaneous estimation of both the degrading PSF and the real high-resolution intensity distribution.

Results. We demonstrate the method on synthetic intensity data derived from a radiative magnetohydrodynamics (MHD) simulation, where we applied PSF convolution and noise to obtain a realistic synthetic input data set, similar to observational short-exposure observations. Quantitative comparisons using image quality metrics, histograms, and power spectral analysis confirm that the model can reliably reconstruct the original image from the stack of synthetic short-exposure frames. Finally, we applied our method to high-resolution observations from GREGOR and DKIST and compare to state-of-the-art speckle reconstructions and multi-frame blind deconvolutions. Our results demonstrate the ability to reconstruct small-scale solar features that exceed the reconstruction performance of state-of-the-art reconstruction methods. With this approach, we lay the foundation for future spatially varying PSFs.