Interpretation of LSTM Prediction on Solar Flare Eruption

Abstract:

Classification of strong and weak solar flares using the Space-weather HMI Active Region Patches (SHARP) parameters with the Long Short Term Memory (LSTM) architecture has gained success in our recent work. However, the internally recurrent and non-linear structure of the LSTM leads to the difficulty in understanding which parameters are the major and minor precursors of strong flares picked up by the algorithm. We firstly trained an LSTM model with SHARP parameters along the magnetic polarity inversion line, classifying strong and weak first flares of each active region, with the model featuring a Heidke Skill Score (HSS) of 0.631. Next, we conduct a post-hoc analysis of the LSTM classification results, thus providing interpretation of the trained model. A combination of Dynamic Time Warping (DTW) and Principal Component Analysis (PCA) is applied on the input SHARP parameters. This leads to the conclusion that total photospheric magnetic free energy density (TOTPOT) and mean shear angle (MEANSHR) are the primary precursors of the initial strong flare of an active region. Our future plan is to extract interpretable features out of the spatial distribution of each SHARP parameter so as to advance our understanding of the relationship between the magnetic field geometry and machine learning prediction on solar flares.