Physics-guided Deep Neural Network for Extreme Ultraviolet Solar Image Desaturation

Abstract EUV imagers on board satellites may experience saturation, resulting in impaired images. To correct this, deep learning has been employed to perform desaturation and eliminate saturation artifacts. It has been observed that deep learning methods have a superior advantage over traditional image restoration methods. However, previous deep learning models did not effectively utilize the physics-related principles of optical imaging, diffraction, and saturation. In this work, we propose a physics-guided deep learning model that incorporates the imaging and saturation principles of the Solar Dynamics Observatory Atmospheric Imaging Assembly into an optimization objective. The experimental results demonstrate the significant superiority of our proposed model over all previous models. Surprisingly, our model achieves excellent image desaturation with solely physics-related principles without customized deep-learning-specific optimization objectives. This result inspires us to solve a well-defined optimization problem that was previously solved by the regularization methods using deep learning. We do not need to reformulate the existing optimization problem to accommodate deep learning. Here, deep learning acts as an advanced solver for an optimization problem, which is exactly suitable for the scenario of artificial intelligence for science with physical laws embedded in neural networks to achieve both knowledge-inspired and data-driven goals.