Presentation #111.06 in the session Solar Flare Prediction.
In Spring 2024, the Solar Flare Sounding Rocket Campaign will involve near-simultaneous rocket launches of the Hi‑C (EUV) and FOXSI (hard X-ray) instruments to study the fine-scale evolution of solar flare features. Ideally, these instruments should observe the flare during its rising phase and peak. To support this Campaign, and in general to provide a mechanism for nowcasting of significant flare activity on a timescale of tens of minutes, we are conducting a study of data provided by the Atmospheric Imaging Assembly on board the Solar Dynamic Observatory (SDO/AIA). Specifically, we train machine learning models (e.g., neural networks) to predict the occurrence of a flare (at least GOES C-class) from patterns indicating the impending solar activity contained in four-dimensional (x, y, t, λ) datacubes recorded by SDO/AIA. Such patterns could include spatial, temporal, or wavelength-dependent changes, or possibly a combination of two, or even all three, of these domains. In addition, we are searching for patterns in datacubes defined on spatial frequency and/or temporal frequency domains (e.g., (u, v, t, λ) or (u, v, ω, λ)), obtained by computing the spatial and/or temporal Fourier transforms of the SDO/AIA data. Our study aims at finding an appropriate number of visibilities (i.e., Fourier components) needed to accurately predict a solar flare occurrence. Such an approach allows a several-order-of-magnitude reduction in the size of the datacubes involved (e.g., from 4096×4096 spatial pixels to a ~20×20 set of visibility points in the spatial frequency (u, v) plane). It builds on a significant heritage of analysis of solar flare activity from instruments that inherently use visibility data, such as the Reuven Ramaty High Energy Solar Spectroscopic Imager (RHESSI) and the Spectrometer/Telescope for Imaging X-rays (STIX) on-board Solar Orbiter.
We will report on the results of application of machine learning techniques to datacubes in both conventional and Fourier transformed spaces. We will validate the performance of the trained models by evaluating appropriate skill-scores of their predictions on test sets of SDO/AIA data corresponding to both flaring and non-flaring periods. Future work will be devoted to the implementation and the assessment of a pipeline for real-time flare nowcasting from SDO/AIA data.