The European Space Agency (ESA) project DEeply uNderstanding Space weathER (DENSER), led by Space Applications Services NV/SA, investigated how machine learning could improve space weather forecasts and provided recommendations for future developments. The project started with an extensive literature study covering all space weather domains.
Two forecast models were developed for time series predictions for phenomena from two completely different domains:
- the proton flux just outside the magnetosphere due to solar eruptions (ForMaL-SEP)
- the ground based magnetic field north component following geomagnetic disturbances (ForMaL-Xrange).
BIRA-IASB coordinated several tasks and contributed to the development of ForMaL-SEP.
Solar Energetic Particles
Solar flares and coronal mass ejections can accelerate particles like electrons, protons, and heavier ions to high energies and escape into the interplanetary space. The observed sudden particle flux increases of several orders of magnitude can last for several days or even weeks. The Earth’s magnetic field deflects most of these particles and those able to penetrate are predominantly absorbed in the atmosphere.
Damage to avionics, communication disturbances and an increased absorbed radiation dose for aircrew and passengers on polar flight routes, where the particles can penetrate the magnetosphere more easily, are potential adverse consequences. Satellites, spacecraft and astronauts are less protected by Earth’s natural shields and extra caution is required.
BIRA-IASB developed the processing and cleaning techniques for the solar X-ray and proton flux measurements that serve as input to ForMaL-SEP and provided scientific support during the model training. An extensive validation indicated that the current model predicts too many false positives.
The teams involved are currently exploring possibilities to reduce these false positives by improving the model training as well as including additional input data.
An end user test campaign coordinated by BIRA-IASB combined with the results and experience gained from the other project activities provided many recommendations for future improvements to the developed models.
Furthermore, the project also provided recommendations for where and how machine learning techniques could potentially be applied to advance forecast models within the space weather domain, and for infrastructure needs to support such developments.