Artificial intelligence improves prediction of solar storms

In a recent "Space Weather" study, an international team led by the Central Institute for Meteorology and Geodynamics (ZAMG) and the Institute of Space Research (IWF) of the Austrian Academy of Sciences was able to combine established models of the solar wind with new machine-learning algorithms to significantly improve the prediction of space weather.

Solar storm predictions by old (blue) and new (orange) model and the measured values. (ZAMG/NASA)

Space weather not only provides impressive light shows, also known as the aurora borealis, but can also have a considerable impact on our modern technologies. So-called geomagnetic storms, for example, can significantly affect the power supply, GPS systems, and other communications systems on which our modern society depends. Expanding our space programs and increasing human presence in space, such as on the International Space Station or soon again on the Moon, requires accurate prediction of the solar wind.

The solar wind is a stream of charged particles that propagates outward from our central star into space and also encounters the Earth's magnetic field. Predicting such solar storms is an unsolved problem even for the most advanced physical models. In various fields such as finance, logistics or classical meteorology, enormous progress is currently being made with artificial intelligence methods when it comes to predicting future developments. An application to the solar wind is therefore obvious, even if space probes still generate significantly less data overall than in the other areas mentioned.

In the new study, researchers from ZAMG and IWF Graz, in collaboration with the US Air Force Research Lab and NASA Goddard, sought a new approach to improve prediction. "We succeeded in combining models of the sun's magnetic field with algorithms from the field of machine learning to predict the effects of the solar wind on Earth much better than was previously possible," explains ZAMG researcher Rachel Bailey, lead author of the study. "We had to study 25 years of measurements and model data. But the results more than justify this effort," Bailey continues. The new method reduces the error in the solar wind speed from 99 km/s to 78 km/s, while generally improving the prediction by about 20%. "The new approach paves the way for optimizing more elaborate simulations of the solar wind in real time in the future," adds IWF researcher and model developer Martin Reiss.

The research project is led by the IWF and funded by the FWF.

Publication

R.L. Bailey, M.A. Reiss, C.N. Arge, C. Möstl, M.J. Owens, U.V. Amerstorfer, C.J. Henney, T. Amerstorfer, A.J. Weiss, J. Hinterreiter: Using gradient boosting regression to improve ambient solar wind model predictions, Space Weather, doi: 10.1029/2020SW002673, 2021.