Solar Storms Arrive Before the Forecast

Space weather events — driven by solar flares, coronal mass ejections, and solar wind variability — disrupt GPS/GNSS positioning, HF communications, power grids, and satellite operations. But we cannot reliably forecast ionospheric and thermospheric conditions more than hours in advance, and even with perfect knowledge of solar drivers, our models produce inaccurate thermospheric neutral density predictions because of incomplete understanding of the coupled Sun-magnetosphere-ionosphere-thermosphere system. Models rarely provide uncertainty estimates, making it impossible to assess forecast confidence. The research-to-operations gap is severe: sophisticated research models exist but are not effectively transitioned to NOAA's Space Weather Prediction Center.

The 2003 Halloween solar storms caused a $450M satellite loss, a Swedish power blackout, and aviation disruptions. A Carrington-class event (1859 magnitude) could cause $1–2 trillion in damage to modern infrastructure in the first year, with recovery taking 4–10 years for the power grid. As society becomes more dependent on GPS (precision agriculture, autonomous vehicles, financial timestamping) and satellite communications (Starlink, OneWeb), vulnerability to space weather increases. The current solar cycle (25) is more active than predicted, with the strongest solar storm since 2003 occurring in May 2024.

Even assuming perfect forecast of model drivers (solar wind speed, density, magnetic field), thermospheric density predictions are often inaccurate due to incomplete system knowledge — the problem is not just forecasting solar input but understanding how the coupled system responds. GEM's systematic model comparison "Challenges" have revealed large discrepancies between models and observations, with no consensus on which physical processes dominate. Lower atmosphere forcing (meteorological waves propagating upward) creates significant ionosphere/thermosphere variability that current space weather models largely ignore. Data assimilation can extend forecast skill to 10–15 days during stratospheric sudden warming events, but skill is much shorter during quiet periods. The research-to-operations pipeline fails because research models are too complex or too slow for operational use, and simplified operational models sacrifice the physics needed for accuracy.

Machine learning approaches for thermospheric density prediction with proper uncertainty quantification. Construction of a Geospace General Circulation Model (GGCM) with genuine predictive capability — GEM's ultimate goal. Better observational networks for the thermosphere and ionosphere, including commercial data sources (satellite drag data, GNSS signal monitoring). Improved coupling of lower-atmosphere wave forcing into upper-atmosphere models. NSF and NASA invested $17M+ in six 3-year awards targeting plasma irregularity forecasting, geoeffective solar eruption prediction via ML, and integrated magnetosphere-ionosphere-thermosphere modeling.

A student team could use publicly available GNSS total electron content (TEC) data from IGS (International GNSS Service) and solar wind data from ACE/DSCOVR satellites to build a machine learning model predicting ionospheric disturbances from solar wind conditions, benchmarking against existing operational models. Alternatively, a team could analyze satellite drag data (publicly available from Space-Track.org) to characterize thermospheric density variability during recent geomagnetic storms. Relevant skills: space physics, machine learning, signal processing, data science.