Detecting spatiotemporal correlation in microflares for improved solar eruption forecasting

Abstract:

Solar magnetic eruptions are the root cause of major space weather events, generating solar flares and coronal mass ejections (CMEs) that in turn generate radiation storms and trigger geomagnetic storms at Earth. Historically, solar eruptions have been among the most difficult space weather phenomena to predict – they are impulsively triggered and extremely rapidly developing events with no obvious precursors identified to date. Here we describe a new analysis technique that applies machine learning technology to the multi-petabyte Solar Dynamics Observatory (SDO) database to search for spatiotemporal correlations in both magnetic field changes and in so-called “microflare” events in the solar atmosphere that may precede major eruptions. The applicable analogy is to earthquake swarms that often precede volcanic eruptions: are there certain patterns in the rate and location of microflares relative to the surrounding magnetic field changes that are precursors to larger eruptions? While the complexity and rapidity of the phenomena preclude human-based detection, this kind of pattern recognition problem is ideally suited to deep learning Convolutional Neural Network (CNN) systems trained on high-cadence SDO data taken in the hours leading up to eruptive events. We employ novel 3D and 4D CNN architectures, adding wavelength and time as additional axes to the traditional 2D spatial pattern analysis architecture. The input is time series of SDO/HMI vector magnetograms and continuum images along with cotemporal SDO/AIA chromospheric, transition region, and coronal images. The target output is an accurate probability of eruption and expected magnitude of associated X-ray flaring within a 3—12-hour forecasting window. This is a newly initiated investigation under NASA’s 2020 Space Weather Operations-to-Research program – the presentation will focus on end-user requirements for flare forecasting, preliminary system architecture designs, and data pre-processing considerations.