Identifying Flux Rope Signatures Using a Deep Neural Network

Abstract

Among the current challenges in space weather, one of the main ones is to forecast the internal magnetic configuration within interplanetary coronal mass ejections (ICMEs). The classification of such an arrangement is essential to predict geomagnetic disturbances. When a monotonic and coherent magnetic configuration is observed, it is associated with the result of a spacecraft crossing a large flux rope with the topology of helical magnetic field lines. This article applies machine learning and a current physical flux rope analytical model to identify and further understand the internal structure of ICMEs. We trained an image recognition artificial neural network with analytical flux rope data, generated from the range of many possible trajectories within a cylindrical (circular and elliptical cross-section) model. The trained network was then evaluated against the observed ICMEs from Wind during 1995–2015.

The methodology developed in this article can classify 84% of simple real cases correctly and has a 76% success rate when extended to a broader set with 5% noise applied, although it does exhibit a bias in favor of positive flux rope classification. As a first step towards a generalizable classification and parameterization tool, these results are promising. With further tuning and refinement, our model presents a strong potential to evolve into a robust tool for identifying flux rope configurations from in situ data.

Appendices

Appendix A: Metrics

$Accuracy$ (Equation 1) is the ratio of the true cases to the total number of instances. $Precision$ (Equation 2) is the fraction of relevant instances among the retrieved instances, while $Recall$ (Equation 3) is the fraction of the total number of relevant instances that were retrieved. Both $Precision$ and $Recall$ need to be taken into account when evaluating the performance of a predictive model. $F_{1}~Score$ (Equation 4) is a well-established measure of the performance of a predictor that considers both $Precision$ (Equation 2) and $Recall$ (Equation 3). Its ideal value is one, and worst value is zero.

$$\begin{gathered} Accuracy = \frac{True~Positive + True~Negative}{Total~Number~of~Instances}. \end{gathered}$$

(1)

$$\begin{gathered} Precision = \frac{True~Positive}{True~Positive + False~Positive}. \end{gathered}$$

(2)

$$\begin{gathered} Recall = \frac{True~Positive}{True~Positive + False~Negative}. \end{gathered}$$

(3)

$$\begin{gathered} F_{1} Score = \frac{2 * Precision * Recall}{Precision + Recall}. \end{gathered}$$

(4)

Appendix B: Complete Classification

Table 4 contains the results for the classification done in all the 353 cases and has the necessary information to compare the classification done in the reference catalog and the classification done by the DCNN model with different amounts of noise. The events marked with an asterisk (*) were used in the evaluation subset part of the training.

Table 4 List of all the ICMEs from Nieves-Chinchilla et al. (2019). For each selected ICME the table presents, event number, ICME start date, label assigned by the reference catalog, classification with the no-noise model, classification with $5\%$ noise model, and classification with $10\%$ noise model. The events marked with the asterisk (*) were used in the evaluation part of the training.