Reconstruction of an Independent Data-Driven TEC Model Using Machine Learning

This article proposes a new model based on supervised machine learning designed for global total electron content (TEC) prediction without relying on atmospheric or solar parameters. The model uses a feedforward neural network (FFNN) with two hidden layers, giving it low complexity and computational cost. By leveraging machine-learning techniques, this model improves a previously established data-driven model proposed by the authors. Our model is trained using TEC data from solar cycle 23, solar cycle 24, and different combinations of both solar cycles. The model is then tested with global ionospheric maps from the $25\text{th}$ solar cycle, which were obtained from the International GNSS Service (IGS) database. Our model is also tested with TEC data from the Madrigal database over specific locations and on days with different solar activity levels. The International Reference Ionosphere (IRI) model was used as a benchmark to our model throughout these tests. The results prove that training with data from concatenated solar cycles yields the best performance. When tested with IGS data, our model achieved an average mean absolute error (MAE) of $5.33$ TEC units, which is nearly 15.5% less than what IRI achieved. When compared with data from Madrigal, the model achieved an average MAE of 3.9, 7.1, and 19.9 TEC units on days with quiet, active, and extreme solar activities, respectively. In contrast, the IRI model achieved an average MAE of 5.4, 8, and 15.5 for the same days. Remarkably, our new model has a size of only 36 $\mathrm{k}$B, representing over a 1800-fold reduction in size compared to the original data-driven model. Consequently, our proposed model can be regarded as a simple and robust yet precise and independent global TEC model.