Space weather: a dispersion model provides more accurate predictions of southern ionospheric convection during storms

Trained on five years’ worth of SuperDARN and DSCOVR data, three deep neural networks were compared: the probabilistic model performs better as the time horizon lengthens and the event intensifies.

Disruptions caused by the solar wind can degrade satellite navigation, cut radio links and threaten electrical networks. Predicting the ionosphere's response to high latitudes remains difficult: empirical models smooth its nonlinear dynamics, and magnetohydrodynamic simulations are too heavy for real time.

Researchers have trained three architectures to predict complete maps of the electrical potential of the Southern Hemisphere from SuperDARN radars: a diffusion model conditioned by solar wind and interplanetary magnetic field measurements at the Lagrange L1 point (DSCOVR satellite), an unconditioned variant, and a deterministic U-Net. Over five years of data (2020-2025), then on self-regressive runs of up to 350 images and on the March 2015 St Patrick's storm, absent from training, the U-Net remains competitive in the short term and in calm weather. But the advantage of the diffusion model grows with the horizon and intensity of the event, making it, according to the authors, the most promising track for an operational forecast.