NARX Neural Network Derivations of the Outer Boundary Radiation Belt Electron Flux

• Published in: Space Weather 2022-05, Vol.20 (5), p.n/a
• Main Authors: Landis, D. A., Saikin, A. A., Zhelavskaya, I., Drozdov, A. Y., Aseev, N., Shprits, Y. Y., Pfitzer, M. F., Smirnov, A. G.
• Format: Article
• Language: English
• Subjects: Algorithms; Boundary conditions; Charged particles; Computer simulation; Dynamic pressure; Electron flux; Electron radiation; Empirical models; Fluctuations; forecasting (1922, 4315, 7924, 7964); Geomagnetic activity; Geosynchronous orbits; Hindcasting; Machine learning; Modelling; Neural networks; Outer radiation belt; Protons; Radiation; Radiation belts; Radiation effects; Satellite observation; Satellites; Scaling factors; Solar cycle; Solar wind; Solar wind velocity; Wind speed; Wind velocities
• ISSN: 1542-7390; 1539-4964; 1542-7390
• DOI: 10.1029/2021SW002774

We present two new empirical models of radiation belt electron flux at geostationary orbit. GOES‐15 measurements of 0.8 MeV electrons were used to train a Nonlinear Autoregressive with Exogenous input (NARX) neural network for both modeling GOES‐15 flux values and an upper boundary condition scaling factor (BF). The GOES‐15 flux model utilizes an input and feedback delay of 2 and 2 time steps (i.e., 5 min time steps) with the most efficient number of hidden layers set to 10. Magnetic local time, Dst, Kp, solar wind dynamic pressure, AE, and solar wind velocity were found to perform as predicative indicators of GOES‐15 flux and therefore were used as the exogenous inputs. The NARX‐derived upper boundary condition scaling factor was used in conjunction with the Versatile Electron Radiation Belt (VERB) code to produce reconstructions of the radiation belts during the period of July–November 1990, independent of in‐situ observations. Here, Kp was chosen as the sole exogenous input to be more compatible with the VERB code. This Combined Release and Radiation Effects Satellite‐era reconstruction showcases the potential to use these neural network‐derived boundary conditions as a method of hindcasting the historical radiation belts. This study serves as a companion paper to another recently published study on reconstructing the radiation belts during Solar Cycles 17–24 (Saikin et al., 2021, https://doi.org/10.1029/2020sw002524), for which the results featured in this paper were used. Plain Language Summary Earth's radiation belts are comprised of two highly dynamic regions consisting of very energetic charged particles (protons and electrons). This paper presents two models that predict electron fluxes at geosynchronous orbit (i.e., the outer radiation belt) and create a scaling factor that can be used in simulations of the radiation belt. Both models are derived using satellite measurements of energetic electrons and a neural network‐based machine‐learning algorithm, the Nonlinear Autoregressive with Exogenous input (NARX). Common geomagnetic activity indices are used as driving inputs for the model. We compare our geosynchronous electron flux model to satellite observations to showcase their performance. Using our NARX‐derived scaling factor, we reconstruct the radiation belts between July and November 1990, and compare it with contemporaneous satellite measurements to show how our model can reproduce observations. Our model allows us to simulate the historical