Modeling ionograms with deep neural networks: Recent updates

We used deep neural networks to forecast ionograms for different solar activity periods and database sizes. To estimate foF2, we utilized two distinct models that identified the last frequency of each ionogram, as the neural network extrapolated virtual heights for all given frequencies. We optimized the hyperparameters of these models and compared their accuracy with the estimates obtained from the IRI and Sami2 models. Furthermore, we explored the temporal variability of the hyperparameter values by training consecutive datasets and observed how the results changed over time. In this work, we will present our results for a climatological model built using all available data and a nowcast model that can update its coefficients using the most recent data.