Can Machine Learning Warn Us of Approaching Geomagnetic Storms?

Geomagnetic storms — disturbances in Earth’s protective magnetic shield caused by oncoming solar particles — can have real-world consequences. A recent research article explores how machine learning can be used to create an early warning system for these events.

Geomagnetic Storms on the Horizon

white-light image of a coronal mass ejection

A white-light image of a coronal mass ejection taken by the Large Angle and Spectrometric Coronagraph. An extreme-ultraviolet image of the Sun is placed at the Sun’s location. [SOHO/LASCO, SOHO/EIT (ESA & NASA)]

A spacecraft at a distant vantage point glimpses a tangled mass of plasma and magnetic fields emerging from the Sun — a coronal mass ejection — headed our way. It’ll be hours or days before the coronal mass ejection collides with Earth, potentially disrupting radio communications, damaging spacecraft electronics, and threatening power grids. How can we predict if a coronal mass ejection will cause these disastrous consequences?

In a recent publication, a team led by Andreea-Clara Pricopi (Technical University of Cluj-Napoca, Romania) tested the ability of machine learning to predict whether a coronal mass ejection will disrupt Earth’s magnetic shield. This technique may provide a way to anticipate geomagnetic storms days in advance.

An Expansive Sample

Machine learning is a relatively new technique in which computers are trained on a set of inputs with known outcomes. The trained computer can then predict the outcomes of a fresh set of inputs.

Pricopi and collaborators took as inputs the speed, angle, and acceleration of coronal mass ejections identified in white-light images, as well as a measure of the overall solar flare activity. The corresponding output is a measure of how disrupted Earth’s magnetic field became, known as the disturbance storm time index. The team trained the model on these inputs and outputs for a subset of 24,403 coronal mass ejections observed between 1996 and 2014, 172 (0.7%) of which caused geomagnetic storms.

illustration of solar particles impinging on the magnetosphere

Artist’s impression of solar particles interacting with Earth’s protective magnetic shield, or magnetosphere, causing a geomagnetic storm. [NASA]

Because so few of the coronal mass ejections in the sample caused geomagnetic storms, Pricopi and collaborators had to be careful about assessing the model’s performance — after all, a model that simply labeled all 24,403 events as not causing a storm would be 99.3% accurate, but it would be useless as a predictor of geomagnetic storms! The team also wanted to be sure that their model correctly predicted all or most storms, even at the risk of false alarms, since the consequences of failing to prepare for a damaging geomagnetic storm are worse than preparing for a storm that never comes.

Prioritizing Powerful Events

Pricopi and coauthors trained their models on 80% of the data set, reserving the remaining 20% for testing the models’ performance. In order to push the models to prioritize finding geomagnetic storms, the team tested several strategies, including penalizing models that misclassified these events and creating synthetic storms based on real data to bulk up the sample size.

visualization of the model output

A visualization of the best model’s performance on the 20% of the data set reserved for testing. The color of the symbols indicates whether the model result was a true negative (TN, blue), false positive (FP, yellow), true positive (TP, green), or false negative (FN, red). Click to enlarge. [Pricopi et al. 2022]

The best model correctly predicted about 80% of storms. The storms overlooked by the model tended to have poor quality data, and false alarms were most common for certain types of coronal mass ejections, giving clues as to how the model might be improved in the future.

These results show that machine learning can be used to predict geomagnetic storms days in advance using a limited number of inputs. However, the authors acknowledge that models that incorporate data from later in a coronal mass ejection’s evolution are more accurate. This suggests that the technique described in this work could be used to flag potentially damaging events, passing them to more precise models to get more information and improve our ability to prepare for an oncoming storm.

Citation

“Predicting the Geoeffectiveness of CMEs Using Machine Learning,” Andreea-Clara Pricopi et al 2022 ApJ 934 176. doi:10.3847/1538-4357/ac7962