Leveraging Topological Data Analysis and Deep Learning for Solar Flare Prediction

We report on the progress of a novel solar eruption prediction system based on feature engineering using topological data analysis and computational geometry of sunspot magnetogram data as inputs to a deep learning neural network architecture. On a 2D magnetogram of an active region, sunspot active regions manifest as magnetic entities of positive and negative polarity structures. The evolution of the magnetic flux, size and relative distance of these structures play a significant role in the flaring process. We use persistent homology as a means to extract these structures at different spatial scales, and with a computational geometric approach, derive various sunspot features based on strongly interacting positive and negative polarities. Using a dense neural network model, we evaluate the prediction performance of our proposed topological features by comparing them against active region features available with the Solar Dynamics Observatory Helioseismic and Magnetic Imager (HMI) data. As a second test, we augment our features to the much more comprehensive set of ~80 features used by the DeepFlareNet neural network model developed at the National Institute of Information and Communications Technology (NICT), Japan to determine if they provide an improvement in the DeepFlareNet prediction skill score. We compare the performance of the dense neural network models with variants of convolutional neural network (CNN) models trained on magnetogram images of active regions. Finally, we propose changes to the data labeling methodology. Moving away from the traditional approach of labeling single magnetograms as producing solar eruptions in the next 24 hours or not, we evaluate the skill score of our models in the prediction of eruptions using time series inputs for prediction at various time instances in the future.