< Back to previous page

Project

Deep learning prediction and hindsight of flare initiation - DELPHI (DELPHI)

Solar flares are explosive energy release events in active regions on the Sun that emit extensive radiation in a very broad range of wavelengths. Through the acceleration of high energy particles and emission of strong radiation, they have a significant impact on Earth. They can present a radiation hazard for astronauts and pilots, they can damage spacecraft and satellites electronics, they can affect radio communications and GPS signals. But the precise mechanisms that lead to the build-up of free energy in the solar corona and that trigger its release during flares events are still unknown. The realisation of machine learning (ML) techniques, with emphasis on Convolutional Neural Networks (CNN), for flare analysis and prediction will (i) help the Royal / Observatory of Belgium better fulfill their duty as a space weather service, by significantly improving the existing predictive capabilities; and (ii) strengthen our knowledge on the physics of flares. The CNNs, indeed, will not be used as black boxes: we will investigate their “attention” with visualization methods to highlight the importance of specific patterns to the prediction and interpret what they “see” in terms of physical processes. Additionally, images in different wavelengths (heights in the solar atmosphere) will be given as input. During this project, we will implement and evaluate different CNN architectures of growing complexity, to find which are optimal for flare predictions (with or without working in operational conditions), and which provide the best results in improving our understanding of the flare initiation mechanisms. In a second step, we will include information about the time evolution of the active regions (e.g. with Recurrent CNNs) and investigate the use of Deep Learning combining heterogeneous data (time series, multiple wavelength images). We will compare the performances and capabilities of different ML strategies, and analyse the connection between the features detected by the neural networks and flare initiation mechanisms. The project will also publish a report synthesising all the results to provide recommendations for the design and development of an operational tool for flare prediction based on ML, as well as useful insights for the forecasters. Finally, our codes and trained models will be made publicly available.

Date:15 Mar 2020 →  Today
Keywords:Solar flares, strong radiation
Disciplines:Stellar astrophysics, Nuclear astrophysics, High energy astrophysics, astroparticle physics and cosmic rays