Optimizing AI to Forecast Dangerous Space Weather

Space weather describes the variable conditions in near-Earth space, driven by the interaction between the continuous outflow of the Solar atmosphere (the solar wind) and the Earth. Space weather is often benign, representing a steady state to which our infrastructure is well designed. However, during global intense events known as geomagnetic storms the near-Earth space environment becomes energized: these hazardous conditions can damage or destroy key infrastructure such as satellites and power networks.

Space weather forecasting is a relatively young field, yet one that has advanced significantly in the last decade, particularly with the adoption of AI methods. Detailed evaluation of the "first wave" of models has highlighted where we urgently need to improve our capabilities. This project will target one (or more) key avenues of enquiry, to enable the next-generation of space weather forecasting models that we require, for example:

A) Damaging space weather events are rare, with major (but localised) damage occurring a few times a decade. However, this raises a key issue for our ability to train advanced AI models. Whilst we have several decades of data available, from ground observatories and satellites, the events we need to forecast appear infrequently. This manifests as a severe data imbalance. How do we then best produce accurate forecasts, minimizing false alarms yet providing the warning we require?

B) Near-Earth space reacts to the solar wind on a huge range of timescales. Even fast changes in the character of the solar wind can initiate both immediate (~minutes) and delayed (~hours) consequences. Forecasting models need to be provided with an input that summarizes recent conditions, but what volume of historical input do we need to provide?

If the time interval is extensive, requiring a very large input vector, can we reduce the dimensionality of this input while retaining the key information? There are numerous parameters that we can use to describe the solar wind, but which ones provide the most benefit to a multivariate forecasting models?