Shape-based Feature Engineering for Solar Flare Prediction

Varad Deshmukh, Thomas Berger, James D. Meiss, Elizabeth BradleyAAAI 2021 · DOI

Abstract

Solar flare prediction is critical for mitigating the impact of space weather on Earth's infrastructure. Traditional forecasting methods rely on physics-based features or deep learning models that often lack interpretability. This paper introduces a novel approach using shape-based feature engineering, specifically computational geometry and topological data analysis (TDA), to quantify the structural complexity of solar magnetic fields. By extracting Interaction Factors and Betti numbers from magnetograms, we demonstrate that topological signatures are powerful predictors of major flares, outperforming industry-standard physics features and achieving a recall of 0.90.

Introduction

Solar flares are massive eruptions of radiation caused by the sudden reconfiguration of magnetic fields in the Sun's atmosphere. These events can disrupt satellite communications, GPS signals, and power grids on Earth. Current forecasting systems primarily utilize human-expert classifications or machine learning models trained on handcrafted physics-based attributes, such as magnetic field strength and current helicity. However, these features often fail to capture the intricate structural precursors of flaring events. This research explores whether the abstract mathematical "shape" of magnetic regions can serve as a more effective proxy for the underlying physical instability that leads to eruptions.

Magnetogram snapshots showing various solar active regions with differing magnetic complexities.

Methodology

The core of this approach is the transition from traditional physics to Shape-based Feature Engineering. The researchers extract two primary types of features from Spaceweather HMI Active Region Patches (SHARPs) magnetograms:

  1. Computational Geometry: The model identifies clusters of positive and negative magnetic polarity. It then calculates an "Interaction Factor" (IF) based on the magnetic flux of these clusters and the spatial distance between them, quantifying the magnetic stress within the region.
  2. Computational Topology (TDA): Using Persistent Homology, the researchers track the evolution of the magnetic field's shape across varying flux thresholds. They specifically count Betti numbers: $\beta_0$ (the number of connected components) and $\beta_1$ (the number of loops or "holes"). These topological signatures provide a mathematically rigorous way to describe the tangled nature of magnetic field lines.
Segmentation of magnetic polarity clusters (b, c) derived from raw magnetograms (a) to facilitate geometric and topological feature extraction.

These features are subsequently processed by a Multi-layer Perceptron (MLP) to predict the occurrence of M-class or X-class flares within a 24-hour window.

Data and Evaluation

The study utilized a dataset of 438,539 magnetograms from 3,691 active regions (2010–2016). Due to the extreme class imbalance (78:1 ratio of non-flaring to flaring instances), the True Skill Statistic (TSS) was employed as the primary performance metric. The model was tuned to prioritize recall, ensuring that potentially catastrophic major flares were not missed, even at the cost of a higher false alarm rate.

Results and Discussion

The experimental results indicate that shape-based features provide a significant advantage over traditional methods:

  • Traditional SHARP features: TSS = 0.70
  • Geometry features only: TSS = 0.71
  • Topology features only: TSS = 0.75
  • Combined (SHARP + Topology): TSS = 0.75

The topological features alone outperformed the industry-standard physics features by approximately 5%. Most notably, the model achieved a recall of 0.90, successfully identifying 90% of major flares in the test set. This suggests that the abstract mathematical complexity of the magnetic field's shape captures critical instability information that traditional flux measurements might overlook.

Conclusion

This research demonstrates that computational topology and geometry offer a powerful, interpretable alternative to traditional physics-based features in solar flare prediction. While the current model relies on single-snapshot data, future work will involve using Long Short-Term Memory (LSTM) networks to incorporate temporal dynamics, potentially uncovering even more predictive power in the evolution of these magnetic shapes. Balancing precision and recall remains a priority for operational deployment, but the high recall achieved here marks a significant step forward in early warning systems for space weather.

Summary updated

BibTeX
@inproceedings{deshmukh2021shape,
  author = {Varad Deshmukh and Thomas Berger and James D. Meiss and Elizabeth Bradley},
  title = {{Shape-based Feature Engineering for Solar Flare Prediction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i17.17795},
}
Figure