Space Physics Seminar spring-2023

The prediction of solar magnetic eruptions, the root cause of solar flares, coronal mass ejections, and radiation storms, and hence the drivers of all severe space weather impacts, remains one of the highest priority challenges in space weather forecasting. Currently, manual analysis of sunspot images and magnetograms followed by climatology prediction and forecaster-in-the-loop judgements is the state of the art. NASA’s Solar Dynamics Observatory (SDO) has now accumulated over 20 PB of photospheric magnetic field, chromospheric, and coronal images over Solar Cycle 24, enabling the development of machine learning (ML) solar flare prediction models and holding out hope that these models will achieve the sort of breakthroughs in skill demonstrated by ML models in other fields such as image classification. In this talk, I discuss three stages of development of ML solar flare prediction models at the CU Space Weather Technology, Research, and Education Center (SWx TREC): topological data analysis of magnetogram data, Convolutional Neural Network (CNN) analysis of magnetogram images, and Self-Attention Network (SAN) analysis of Extreme Ultraviolet (EUV) chromospheric and coronal images. Our models perform comparably to other ML flare prediction models as well as to the state-of-the-art manual methods, and we demonstrate conclusively that pattern recognition can predict eruptions as skillfully as physics-based methods. But our models, like all other ML flare predictions models to date, fail to achieve breakthrough performance in predictive skill. We discuss possible reasons for this as well as several common errors that lead to deceptively high skill scores occasionally reported in the literature.