An automated classification approach to ranking photospheric proxies of magnetic energy build-up

TL;DR

This study uses machine learning on photospheric magnetic field data from ~2000 active regions to predict solar flares, achieving moderate skill scores and highlighting key magnetic features related to energy build-up.

Contribution

It introduces a comprehensive classification approach combining multiple magnetic parameters and wavelet analysis for flare prediction, revealing the limits of current large-scale observational methods.

Findings

True skill score of ~0.5 for 2-24hr prediction windows

Magnetic gradient features are most predictive

Photospheric magnetic field alone is insufficient for precise flare forecasting

Abstract

We study the photospheric magnetic field of ~2000 active regions in solar cycle 23 to search for parameters indicative of energy build-up and subsequent release as a solar flare. We extract three sets of parameters: snapshots in space and time- total flux, magnetic gradients, and neutral lines; evolution in time- flux evolution; structures at multiple size scales- wavelet analysis. This combines pattern recognition and classification techniques via a relevance vector machine to determine whether a region will flare. We consider classification performance using all 38 extracted features and several feature subsets. Classification performance is quantified using both the true positive rate and the true negative rate. Additionally, we compute the true skill score which provides an equal weighting to true positive rate and true negative rate and the Heidke skill score to allow comparison to other flare forecasting work. We obtain a true skill score of ~0.5 for any predictive time window in the range 2-24hr, with a TPR of ~0.8 and a TNR of ~0.7. These values do not appear to depend on the time window, although the Heidke skill score (<0.5) does. Features relating to snapshots of the distribution of magnetic gradients show the best predictive ability over all predictive time windows. Other gradient-related features and the instantaneous power at various wavelet scales also feature in the top five ranked features in predictive power. While the photospheric magnetic field governs the coronal non-potentiality (and likelihood of flaring), photospheric magnetic field alone is not sufficient to determine this uniquely. Furthermore we are only measuring proxies of the magnetic energy build up. We still lack observational details on why energy is released at any particular point in time. We may have discovered the natural limit of the accuracy of flare predictions from these large scale studies.