Self-attention networks for solar flare prediction using SDO/AIA EUV data and a novel AIA-based flare catalog

There is growing interest in using ultraviolet and extreme ultraviolet images from instruments such as SDO/AIA for solar flare prediction as these images may reveal more features associated with flaring than photospheric magnetic field data alone, which has been the primary data source to date for machine-learning based forecasting. Along the way to developing a probabilistic solar flare forecast using these data we seek to understand how to define flaring activity given only AIA data, thus allowing us to move away from the GOES single pixel X-ray intensity measurement definition of solar flares. We demonstrate that we are able to estimate flare magnitudes with high accuracy using an extremely randomized trees regression model whose inputs are scalar features extracted from the AIA image cutouts of flares. To generate an AIA-based flare catalog, we combine these magnitudes with flare start, peak, and end times defined by a novel peak finding algorithm applied to AIA time series data. Using this catalog to label our machine learning dataset, we investigate self-attention based networks for solar flare prediction with temporal stacks of AIA image cutouts as input. Self-attention based machine learning architectures, such as transformers, have been used traditionally for natural language processing and are increasingly being used for image-based tasks with high success.