Optimizing the neural network modeling of ionospheric F2-peak parameters

The peak of the F2 layer in the ionosphere is a crucial anchor point in many electron density modeling methods. It is essential to predict the peak of the F2 layer accurately in order to create reliable models of the ionosphere. Recently, machine learning approaches have shown excellent results in predicting electron density, often surpassing the traditional empirical models of the ionosphere in terms of accuracy. 

In this presentation, we analyze the neural network-based model of electron density in the topside ionosphere (NET) and optimize the hyperparameters of NET's submodels for NmF2 and hmF2. The dataset used in this study consists of radio occultation (RO) observations from the CHAMP, GRACE, and COSMIC-1 satellite missions from 2001 to 2019. The inputs to the submodels include geomagnetic latitude and longitude, universal time, day of the year, and the P10.7, Kp, and SYM/H indices. The tuned parameters in the hyperparameter optimization (HPO) were the sizes of each of the three hidden layers, activation function, dropout rate, standard deviation of the regularizing Gaussian noise layers, orders of the Fourier features (FFT) for periodic inputs, necessary number of Kp index observations, learning rate, and batch size.  

We analyze the effects of regularization on the performance of both submodels, and find the optimal values that balance the bias-variance tradeoff. We also perform the feature selection and show that the history of the Kp index of up to 15 hours is important for reproducing the ionospheric behavior, which is in line with known physical evolution of the ionosphere during geomagnetic storms. The optimized models reproduce the effects of several physical processes, including complex dynamics driven by neutral winds and electromagnetic drifts. We showcase physical features depicted by the NET model and interpret them in combination with in-situ measurements of the plasma drifts.