Machine Learning for Real-Time Solar Flare Detection, FHNW School of Computer Science

Whether in space research, industrial production or other data-intensive environments, the challenge is often the same: large amounts of complex sensor data need to be analysed quickly and transformed into useful information.

Solar research provides a particularly demanding example. Solar flares are among the most energetic events in our solar system. Understanding where they occur and how they develop helps researchers investigate the physical processes taking place on the Sun.

The Spectrometer/Telescope for Imaging X-rays (STIX) aboard ESA’s Solar Orbiter provides valuable measurements of solar flares. However, STIX does not capture conventional images. Instead, images and the positions of solar flares have to be reconstructed from complex measurement data.

This project investigates how Machine Learning and Deep Learning can make this process significantly faster.

The project team enhanced an existing Machine Learning model that estimates the position of a solar flare directly from STIX measurement data. The model was trained and evaluated using observational data and optimised to cover a larger area of the Sun.

The localisation method was then combined with a Deep Learning approach for image reconstruction. Together, the two methods create an automated workflow from complex detector data to flare localisation and a reconstructed image.

The developed solution has been integrated into the STIX Data Center, bringing the Machine Learning models directly into the researchers’ workflow.

Researchers can select a time and energy range and receive both an estimated position of the solar flare and a reconstructed quick-look image in near real time.

On independent observational data, the Machine Learning model achieves a mean positional error of around 22 arcseconds. At the same time, it determines the flare position around 60 times faster than traditional methods and covers a larger area of the Sun than the previous solution.

The methods demonstrated in this project are not limited to astrophysics. Many organisations face similar challenges: large volumes of sensor, measurement or image data need to be processed efficiently, patterns identified automatically and results made available quickly for decision-making or further analysis.

Our expertise in Machine Learning, Deep Learning, image reconstruction, sensor data analysis and AI-driven data processing can be transferred to applications in industry and other fields.

Together with companies and research partners, we develop AI solutions for complex data challenges – from feasibility studies and prototypes to applied research and integration into real-world processes.