A Hybrid CNN–Transformer Network to EnhanceSolar Magnetogram Resolution for Flare PredictiveAnalytics

The disparity in spatial resolution between the SOHO/MDI and SDO/HMI magnetogramscreates inconsistencies that hinder reliable long term studies of solar magnetic fields and reduce flareforecasting. To solve this problem, the MagRes-Net a hybrid super-resolution Network that uses convolutionalneural networks (CNNs) to extract features locally and transformer-based self attention toextract global gradients. The designed network reconstructs high-resolution magnetograms from lowerresolution images while maintaining physical structures realistic features of the magnetic field. The modelis trained using MDI-HMI co-aligned image pairs and implemented physics-aware constraints into themodel, such as gradient consistency, and magnetic flux conservation. Through quantitative evaluations,MagRes-Net was better than interpolation and CNN-based methods in terms of peak signal-to-noise ratio (PSNR), mean squared error (RMSE), Learned Perceptual Image Patch Similarity (LPIPS) andstructural similarity index measure (SSIM). Quantitatively, AR coverage increases from 12.23% in nativeMDI observations to 13.25% in the super-resolved output, approaching the SDO/HMI reference valueof 14.48%. it indicates the model successfully recovers resolution-dependent attributes. Additionally,spectral, noise robustness and perceptual analyses demonstrated that MagRes-Net preserves structureat all frequencies, and thus can improve resolution dependent magnetic diagnostic techniques, including;continuity of polarity inversion lines, morphology of strong field regions, and magnetic flux distribution.Therefore, these results demonstrate that the proposed method bridges the cross-instrument resolutiondisparities, allowing for the creation of homogeneous magnetogram data sets that are capable for longterm studies relevant for solar activity and improved flare forecasting.