Electric Potential Patterns Forecasting in the Southern Hemisphere with Deep Learning Techniques

Space weather disturbances driven by the solar wind can degrade satellite navigation, disrupt radio communications, and threaten power infrastructure, making accurate forecasting of the high-latitude ionosphere's response a critical operational need. Existing approaches, empirical climatological models and physics-based magnetohydrodynamic simulations, either smooth out the ionosphere's time-dependent, non-linear response or are too computationally expensive for real-time use, while prior machine learning efforts have mostly targeted scalar indices rather than the full spatial structure of ionospheric convection. Here we train and compare three deep learning architectures, a probabilistic diffusion model conditioned on multi-variate solar wind and interplanetary magnetic field measurements at the L1 Lagrangian point, an unconditioned diffusion ablation, and a deterministic U-Net baseline, to forecast Southern Hemisphere high-latitude electric potential maps derived from SuperDARN radar observations. Using five years (2020–2025) of SuperDARN data synchronised with DSCOVR L1 measurements, we evaluate the models under a single-pass regime, an extended autoregressive rollout of up to 350 frames, and an out-of-distribution case study on the intense March 2015 St. Patrick's Day storm, unseen during training. We find that the relative advantage of the deterministic and probabilistic approaches is not fixed: the deterministic model is competitive over short horizons and calm conditions, while the diffusion model's advantage grows and eventually dominates as the forecast horizon lengthens and the event becomes more dynamic, evidence that probabilistic, sample-based generative models are the more promising direction for operational, long-horizon space weather forecasting.

• Open, paired dataset: five years (2020–2025) of Southern Hemisphere SuperDARN electric potential maps, already time-matched to DSCOVR L1 solar wind/IMF measurements and deduplicated (one L1 sample per map), released with the code — no need to rebuild the L1-to-map pairing pipeline from scratch. See the code repository.
• Probabilistic method: a diffusion-based generative model conditioned on multi-variate L1 solar wind/IMF drivers, producing an ensemble of physically plausible convection maps per forecast — unlike prior SuperDARN deep learning models, which are all deterministic point estimators with no notion of forecast uncertainty.
• Evaluation pushed to the limit on two axes: (i) horizon — from a single-pass regime matching the training setup to an extended autoregressive rollout of up to 350 predicted frames; (ii) distribution shift — a fully out-of-distribution case study on the intense March 2015 storm, which is out-of-distribution in two independent ways at once: it belongs to Solar Cycle 24 rather than the Solar Cycle 25 interval the models were trained on, and its L1 driver comes from ACE rather than DSCOVR, the instrument used throughout training — so the model must generalise across both a different storm regime and a different measuring instrument simultaneously.

• The model forecasts the ionosphere's response given an observed L1 solar wind/IMF driver — it does not forecast the driver itself. Every horizon reported here is still conditioned on the true, measured L1 timeseries, so this is not yet a fully autonomous forecast.
• Closing that gap needs an upstream model that predicts L1 conditions directly from solar-disk observations, removing the dependence on a real-time L1 measurement and enabling a genuine end-to-end Sun-to-Earth pipeline. That is the natural next step, not something attempted here.
• This study covers the Southern Hemisphere only; extending to the Northern Hemisphere would further test how broadly the results hold.