Application and interpretation of deep learning for identifying pre-emergence magnetic-field patterns

TL;DR

This paper applies deep convolutional neural networks to classify pre-emergence solar active regions from magnetogram data, achieving high accuracy and providing interpretability of the features involved.

Contribution

It introduces a CNN-based method for identifying pre-emergence magnetic patterns and develops a pruning algorithm for interpreting the neural network's learned features.

Findings

CNN classifies pre-emergence magnetograms with ~85% TSS 3 hours prior to emergence

CNN outperforms baseline methods based on unsigned magnetic flux

Network interpretation reveals sensitivity to magnetic region scale and intensity

Abstract

Magnetic flux generated within the solar interior emerges to the surface, forming active regions (ARs) and sunspots. Flux emergence may trigger explosive events, such as flares and coronal mass ejections and therefore understanding emergence is useful for space-weather forecasting. Evidence of any pre-emergence signatures will also shed light on sub-surface processes responsible for emergence. In this paper, we present a first analysis of emerging ARs from the Solar Dynamics Observatory/Helioseismic Emerging Active Regions (SDO/HEAR) dataset (Schunker et al. 2016) using deep convolutional neural networks (CNN) to characterize pre-emergence surface magnetic-field properties. The trained CNN classifies between pre-emergence (PE) line-of-sight magnetograms and a control set of non-emergence (NE) magnetograms with a True Skill Statistic (TSS) score of ~85%, 3h prior to emergence and ~40\%, 24h prior to emergence. Our results are better than a baseline classification TSS obtained using discriminant analysis of only the unsigned magnetic flux. We develop a network pruning algorithm to interpret the trained CNN and show that the CNN incorporates filters that respond positively as well as negatively to the unsigned magnetic flux of the magnetograms. Using synthetic magnetograms, we demonstrate that the CNN output is sensitive to the length-scale of the magnetic regions with small-scale and intense fields producing maximum CNN output and possibly a characteristic pre-emergence pattern. Given increasing popularity of deep learning, techniques developed here for interpretation of the trained CNN -- using network pruning and synthetic data -- are relevant for future applications in solar and astrophysical data analysis.