SHINE: Exploring the Initiations of Solar Flares using Deep Learning Methods

Many new discoveries of phenomena of solar activity have been made in recent years thanks to state-of-the-art instrumentation for both ground-based and space-borne observations. However, due to ever increasing spatial and temporal resolutions, researchers are facing tremendous challenges to handle massive amounts of data in near real-time, and to extract important information from the data that can lead to further scientific discoveries and forecasting of solar activities. These tasks become more demanding when larger telescopes are revealing finer structure with rapid dynamics and evolution, such as the NSF-funded 1.6 meter Goode Solar Telescope (GST) at Big Bear Solar Observatory (BBSO). This project addresses the Solar, Heliospheric, and Interplanetary Environment (SHINE) goal of understanding the initiation of solar flares through use of deep learning methods and solar observations from the GST. The project is a joint effort between physics and computer science groups, integrating research and education through interdisciplinary training. A graduate researcher and high school students will be trained.In this research project, the team will develop and apply a suite of deep learning models and tools to advance the understanding of the initiation of solar flares, and provide near real-time flare forecasting. There are five interrelated tasks. (1) They will develop a convolutional neural network with attention mechanisms for inverting GST Stokes profiles to vector magnetograms with high efficiency and reduced noise. (2) Using a Bayesian convolutional network with uncertainty quantification, they will trace the fibril and loop structures of chromospheric observations to provide an assessment of magnetic fields in the chromosphere, which is crucial in 3D magnetic field extrapolations. (3) They will train generative adversarial networks (GANs) using simultaneous NASA Solar Dynamics Observatory (SDO) and GST magnetograms to create higher-resolution, larger field-of-view data, which are critical to derive flow fields in flare-producing solar active regions. (4) They will train a new GAN model, using SDO vector magnetograms and Hα images to derive transverse fields from the NASA Solar Heliospheric Observatory line-of-sight magnetograms. Therefore, the availability of vector magnetograms is extended to two solar cycles. (5) They will develop a new encoder-decoder bidirectional long short-term memory network with attention mechanisms to carry out near real-time flare prediction and evaluate the most critical magnetic parameters relevant to flare initiations. With these data and tools, they will address the following two key science questions: (i) With consistent high-resolution observations, what roles do the evolution of magnetic fields and flow fields play in storing energy and triggering solar flares? (ii) What is the quantitative assessment of flare prediction with the deep learning-processed data and deep learning-based prediction tools?This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.