Abstract
Traditional deep learning models for the ionosphere often rely on space weather indices as inputs, offering predictive capability but limited physical interpretability. We present SRON2NN, a physics‐motivated artificial neural network (ANN) that incorporates two geophysically motivated features: effective solar radiation (ESR) and the thermospheric column density ratio of atomic oxygen to molecular nitrogen (ΣO/N 2 ). Using a comprehensive data set comprising COSMIC‐1 radio occultation data, TIMED/Global Ultraviolet Imager ΣO/N 2 observations, and solar EUV measurements from 2008 to 2019, we find that SRON2NN consistently outperforms an index‐based ANN using only Dst, F10.7, time, and location across all testing conditions. The improvement is most evident during geomagnetic storms in mid‐latitude regions, especially in the Southern Hemisphere, where the contribution of ΣO/N 2 leads to an 8.20% reduction in root mean square error. Compared with adding vertical total electron content and IRI‐NmF2, our physics‐motivated features yield a more pronounced improvement in reconstructing NmF2. Cross‐model experiments suggest that incorporating these two features can consistently boost performance not only in ANN frameworks but also in tree‐based models, including random forests and XGBoost. Further analysis of four geomagnetic storm events shows that SRON2NN is generally consistent with storm‐time ΣO/N 2 depletion and the associated decrease in NmF2. Overall, these results suggest that incorporating ESR and ΣO/N 2 as physically motivated inputs can improve NmF2 reconstruction while enhancing interpretability.

This publication has 44 references indexed in Scilit: