Recovering SHARP Magnetic Parameters from Solar Vector Magnetograms with Swin Transformer

Magnetic-field parameters of solar active regions, such as the 16 Space-weather HMI Active Region Patches (SHARP) indices, are widely used to characterize active region magnetic complexity. Here we investigate whether deep-learning models can learn the physical information encoded in vector magnetograms well enough to recover these diagnostics directly from the images, thereby establishing a representation that can support subsequent downstream tasks. We propose a hybrid CNN-Swin Transformer regression framework, where the CNN extracts local magnetic features and the Swin Transformer captures large-scale spatial context through hierarchical, shifted-window self-attention. The model is trained on 750,430 SHARP vector magnetograms from 2011 – 2024 with a loss that combines mean squared error and Pearson correlation coefficient (PCC). On a held-out test set, the model achieves an average PCC of 0.934 across the 16 parameters. Performance is highest for geometry- and shear-related quantities (e.g., MEANGAM, MEANSHR, SHRGT45) and for non-potentiality proxies (e.g., MEANPOT), while current- and derivative-sensitive parameters remain more difficult (e.g., MEANJZD). These results indicate that the proposed model can recover physically meaningful SHARP diagnostics from vector magnetograms and can serve as a foundation for downstream active-region analysis and prediction tasks.