Modeling ionograms and critical plasma frequencies with neural networks

Ionosondes offer broad spatial coverage of the lower ionosphere, supported by a global network of affordable instruments. This motivates the exploration of new methods that exploit this geographical coverage to capture spatially dependent characteristics of electron density distributions using data-driven models. These models must have the versatility to learn from ionogram data. In this work, we used neural networks (NN) to forecast ionograms across two solar activity cycles. The ionosonde data was obtained from the digisonde at the Jicamarca Radio Observatory (JRO). Each NN comprises one NN that estimates the ionogram trace and another one that estimates the critical frequency. Two forecasting models were implemented. The first one was trained with all available data and was optimized for accurate predictions along that time range. The second one was trained using a rolling-window strategy with just 3 months of data to make short-term ionogram predictions. Our results show that both models are comparable and can often outperform predictions by empirical and numerical models. The hyperparameters of both models were optimized using a specialized library. Our results suggest that a few months of data was enough to produce predictions of comparable accuracy to the reference models. We argue that this high accuracy is obtained with short time series because the NN captures the dominant periodic drivers. Finally, we provide suggestions for improving this model.