Solar wind prediction using deep learning

TL;DR

This paper employs deep learning to predict solar wind speed from solar corona images, outperforming traditional models and revealing correlations between coronal features and wind properties.

Contribution

It introduces a novel deep learning approach using EUV images to forecast solar wind, capturing key coronal features associated with wind speed without explicit physics modeling.

Findings

Deep learning model outperforms autoregressive and naive benchmarks.

Model identifies coronal holes as indicators of fast wind.

Coronal features correlate with wind speed predictions.

Abstract

Emanating from the base of the Sun's corona, the solar wind fills the interplanetary medium with a magnetized stream of charged particles whose interaction with the Earth's magnetosphere has space-weather consequences such as geomagnetic storms. Accurately predicting the solar wind through measurements of the spatio-temporally evolving conditions in the solar atmosphere is important but remains an unsolved problem in heliophysics and space-weather research. In this work, we use deep learning for prediction of solar wind (SW) properties. We use Extreme Ultraviolet images of the solar corona from space based observations to predict the SW speed from the NASA OMNIWEB dataset, measured at Lagragian point 1. We evaluate our model against autoregressive and naive models, and find that our model outperforms the benchmark models, obtaining a best-fit correlation of 0.55 $\pm$ 0.03 with the observed data. Upon visualization and investigation of how the model uses data to make predictions, we find higher activation at the coronal holes for fast wind prediction ($\approx$ 3 to 4 days prior to prediction), and at the active regions for slow wind prediction. These trends bear an uncanny similarity to the influence of regions potentially being the sources of fast and slow wind, as reported in literature. This suggests that our model was able to learn some of the salient associations between coronal and solar wind structure without built-in physics knowledge. Such an approach may help us discover hitherto unknown relationships in heliophysics datasets.