Statistical Methods for Interpretable and Trustworthy Solar Flare Prediction

Oregon Convention Center 

The challenge of predicting solar flares has been tackled frequently in recent years, thanks to the burst of high-quality solar data. Among all sources of solar data, the Helioseismic and Magnetic Imaging (HMI) and Atmospheric Imaging Assembly(AIA) data contain abundant information about the Sun in the format of multi-channel images, and applications of this data in deep learning models have been proven to be successful. However, these methods are not interpretable and cannot provide uncertainty quantification (UQ). In this talk, we develop an interpretable and trustworthy statistical method that treats the HMI/AIA data as tensor data. We propose to predict flare intensity with the tensor data in a novel framework called Tensor-GPST, where we first transform the high-dimensional tensor data into a low-dimensional latent tensor via sparse tensor contraction, and then the latent tensor is used for prediction via the Gaussian process. We introduce an anisotropic total-variation regularization when contracting the tensor and estimate the model with alternating proximal gradient descent. We validate our approach via simulation and real application to flare intensity forecasting.