Enhancing Deep Learning Ionospheric Modeling With Solar Radiation and Flare Classes

Abstract

The ionosphere is pivotal for satellite navigation, radio communication, and the modeling of space weather. However, the accurate three-dimensional modeling of ionospheric features remains a challenge. Since solar activity introduces changes in space weather, we collected COSMIC radio occultation observations of 2010–2020 with a suite of indices related to solar and geomagnetic activities, especially including solar EUV and X-ray radiation fluxes, to develop a deep learning model for the global ionospheric electron density. This model, which is called the Solar Flare and Radiation Neural Network (SFRNN) and is based on Embedding, Long Short-Term Memory and fully connected layers, presented excellent performance in reconstructing ionospheric profiles. In this study, 28-min was found to be the best input solar radiation interval for SFRNN with annual RMSEs of 6.24 × 10⁴ to 1.56 × 10⁵ el/cm³. Significantly, during solar flare events, SFRNN had a lower reconstruction error than the former artificial neural network (ANN) model that only uses space weather indices. The most substantial improvement was observed under X-class flares, where SFRNN exhibited a 18.3% lower Root Mean Squared Error than ANN. To further validate the modeling accuracy, electron density profiles derived from Jicamarca incoherent scatter radar (ISR) were used. SFRNN successfully provided profiles with high consistency with the ISR observation in the ionospheric layers. Our modeling results demonstrate that refined solar activity parameters can effectively improve reconstruction performance.