Speaker
Description
In this new ionospheric modeling study, we developed a physics-informed neural network (PINN) to predict key ionospheric peak parameters: foF2, hmF2, and TEC, using various space weather inputs. The model is trained on a globally distributed dataset that combines ground-based ionosonde measurements from the GIRO network with corrected space-based observations from the COSMIC satellites.
To ensure physical consistency, the neural network is constrained by simplified representations of the Chapman layer production function, the Equatorial Ionization Anomaly (EIA) structure, and other fundamental ionospheric physics. We validated our model against independent ionosonde measurements and compared its performance with the empirical IRI-2020 model. Our results show that the PINN achieves significantly improved accuracy, with an average R² value of 0.91 for the peak parameters predictions, outperforming IRI-2020.
Furthermore, we evaluated the model’s response to geomagnetic storm conditions and found that it captures the global storm-time ionospheric behavior, particularly in low-latitude regions. This enhanced performance under disturbed conditions highlights the model’s robustness.
Our physics-informed approach offers a promising tool for real-time prediction of ionospheric parameters critical for HF radio wave propagation, GNSS positioning, and other space weather–sensitive technologies.
