Physical Interpretability of Deep Learning Models for Solar Flare Prediction

Abstract

Solar flares are sudden releases of energy from the Sun that can produce intense electromagnetic radiation and energetic particles, with the potential to disrupt satellite systems, radio communications, and other space-based and technological infrastructure. Accurate solar flare prediction is therefore important for providing timely warnings, while interpretable models are needed to understand the physical information driving predictions and to assess their reliability in operational settings. This study investigates a neural-network-based solar flare prediction model using SHARP magnetic-field parameters and flare-history information, with particular emphasis on model performance, interpretability, and the effects of incorporating physical constraints. The baseline model achieves competitive predictive performance, while temporal deployment on new data results in a small performance decrease. Physics-informed loss functions are shown to alter the model’s classification behavior without substantially reducing overall performance, although the effect depends strongly on the physical relationship imposed. The Pearson MEANGBZ constraint reduces false negatives at the cost of more false positives, whereas the Kendall MEANGAM constraint produces the opposite behavior and substantially more missed flares. These results demonstrate that TSS alone is insufficient to characterize the operational behavior of flare prediction models and that threshold selection and the balance between false positives and false negatives are important considerations. Finally, a parameter study shows substantial redundancy among the input features: a reduced model using only 11 features achieves a TSS of 0.7152, close to the approximately 0.72 of the full baseline. Overall, the results indicate that accurate, interpretable, and operational solar flare prediction can be achieved with fewer input parameters, while physics-informed learning can be used as a constraint to modify model behavior without necessarily compromising predictive performance.