Improving solar storm modeling with machine learning

In the last decade, the prediction of coronal mass ejections (CMEs) and related space weather phenomena at Earth has attracted a lot of attention from scientists all over the world. Despite the community wide efforts to enhance prediction models, accurately forecasting a CME’s arrival time at Earth is still an unsolved issue, along with an inadequate number of false alarms. A possible reason for that might be the limited observational possibilities of coronagraphs that form the basis of most predictions.

With the STEREO-A spacecraft passing the L5 point of the Sun-Earth system in 2020, its heliospheric imagers (HI) provide white-light observations of CMEs from their launch at the Sun up to Earth from changing viewpoints until 2027 when it will reach L4. HI data are perfectly suited to be used as input for CME modeling and prediction tools. These tools are often developed based on HI science data arriving with a delay of some days. HI beacon data, available in near real-time, however, has a lower spatial and temporal resolution and is therefore less qualified to form the basis of accurate predictions.

Figure 1: Steps of HI image processing. Panel a) shows the raw HI image, in which the F-corona is the most prominent feature, panel b) shows the background subtracted image with the CME and the starfield visible. After starfield reduction a clean image with a clear CME signal is visible in panel c) and panel d) shows the according running difference image.


Figure 2: Time-elongation maps produced from science data (upper panel) as well as from near real-time beacon data (lower panel). The vertical stripes are missing data, while the bright tracks show CMEs as they evolve in the interplanetary space [Bauer+ 2021].


In this project, we aim to enhance STEREO's HI beacon data by applying machine learning tools in order to increase the image quality and to overcome low temporal resolution and data gaps. The resulting artificial HI science data should then serve as input to an HI-based CME modeling and prediction tool. In addition to the data quality itself, the extraction of information from these data is challenging. Usually, the elongation of the CME front within HI data is extracted manually leading to large differences depending on the scientist doing the tracking and even between tracks measured from the same person. By applying an automated CME detection and tracking ML method we aim to decrease the prediction error arising from the measurements themselves.

Team
PI: Tanja Amerstorfer
Postdoc: Justin Le Louëdec
PhD student: Maike Bauer
Master's student: N.N.
Duration: 03/2023 – 02/2026
Link to FWF project page: P 36093





Acknowledgements: This research is funded by the Austrian Science Fund (FWF) [P 36093].