Research

While machine learning has led to immense advances in the amount of data that can be analyzed automatically, one often-mentioned criticism is that "there is no physics in ML" or that it is run as a "black box" without knowing which physical parameters influenced the results. We have therefore started focusing on explainable artificial intelligence (XAI), which are techniques that allow us to reconstruct which feature of the input had the largest influence on a result.

Using two XAI techniques, namely Gradient-weighted Class Activation Mapping (Grad-CAM) (Selvaraju et al. 2017) and Expected Gradients (EG) (Erion et al. 2021) we investigated which features of spectra indicate an upcoming flare. We found that Mg II triplet emission, flows, broad, and highly asymmetric spectra are all important for flare prediction. This indicates that the solar atmosphere is undergoing changes before flares.

Mg II spectra with typical pre-flare signatures. The color-coding indicates which spectral features are most important for flare-prediction, red being of highest importance. The method also predicted the spatial location of the future flare emission ~80% of the time. © Universität Bern

Flare prediction is an extremly difficult problem, in fact, it is not known yet if flares are even predictable. While past studies mostly focused on magnetic field data to make predictions (e.g. Leka&Barnes 2003, Bobra & Couvidat 2015), we investigated the use of UV spectra to predict flares. Regions that are about to erupt often show enhanced brightenings and we wanted to test if there were typical spectra that appear only before flares.

We found that we could correctly identify pre-flare spectra approximately 35 minutes before the start of the flare. The accuracy increased to 90% when looking closer in time to the flare. This indicates for the first time that spectra can be used for flare prediction, although more advance warning would be desirable and therefore, we are expanding our models.

Image of a strong flare that occurred on Sep 10, 2014. The right panels (blue and green) show flare spectra. Spectral lines become very bright during flares. © Universität Bern

Solar flares differ by orders of magnitude in energy and duration, but are the basic physical processes the same for all flares? Similar physics would imply similar spectra, in which the information of temperature, density, and velocity of the solar atmosphere is encoded. But comparing millions of spectra manually would be an impossible task.

Machine learning helps us to classify and analyze millions of data points efficiently. We used an adapted version of the k-means classification to investigate 19 flares, which contained several million spectra. We found that the same type of spectra (depicted light and dark blue in the picture left) appear in every single flare, which indicates that all flares seem to have some common physics.

Classification of flare spectra. These 9 types of spectra (right panels) occurred during a flare and their locations on the Sun are color-coded in the left image. © Universität Bern

To observe flares in detail, we need large telescopes with the most modern instrumentation. Our group also specializes in designing astronomical instrumentation, coordinating ground- and space-based observations, and using the best facilities to observe flares.

L. Kleint led the optics redesign of the largest European solar telescope (GREGOR), leading to the highest resolution solar images obtained from Europe to date. An article on Phys.org, as well as an article on medium.com and a scientific publication (Kleint, Berkefeld, Esteves et al, A&A 641, 27, 2020) describe the success.

We also caught the "best-observed" solar flare by coordinating a world-wide observing campaing with several NASA satellites and the Dunn Solar Telescope.

Further publications on our work and on flare physics can be downloaded via the Astrophysics Data System (ADS).

Optical layout of the GREGOR telescope. Courtesy of KIS. © Universität Bern