Prediction of Halo Coronal Mass Ejections Using SDO/HMI Vector Magnetic Data Products and a Transformer Model

TL;DR

This work introduces DeepHalo, a transformer-based framework for predicting halo CMEs using 24-hour SHARP parameter profiles derived from SDO/HMI data. By integrating CME records from DONKI and LASCO (2010–2023) and applying SMOTE-based oversampling to balance the training set, DeepHalo processes 120-time-step profiles with 18 features through a four-encoder, attention-driven architecture, achieving a test TSS of 0.907 and outperforming a comparable LSTM model (0.821). The study demonstrates not only superior predictive performance but also interpretability via average attention heatmaps, revealing when the model focuses on early versus later profile intervals. The results suggest DeepHalo as a feasible, interpretable tool for halo CME forecasting, with potential utility for space weather operations and risk assessment.

Abstract

We present a transformer model, named DeepHalo, to predict the occurrence of halo coronal mass ejections (CMEs). Our model takes as input an active region (AR) and a profile, where the profile contains a time series of data samples in the AR that are collected 24 hours before the beginning of a day, and predicts whether the AR would produce a halo CME during that day. Each data sample contains physical parameters, or features, derived from photospheric vector magnetic field data taken by the Helioseismic and Magnetic Imager (HMI) on board the Solar Dynamics Observatory (SDO). We survey and match CME events in the Space Weather Database Of Notification, Knowledge, Information (DONKI) and Large Angle and Spectrometric Coronagraph (LASCO) CME Catalog, and compile a list of CMEs including halo CMEs and non-halo CMEs associated with ARs in the period between November 2010 and August 2023. We use the information gathered above to build the labels (positive versus negative) of the data samples and profiles at hand, where the labels are needed for machine learning. Experimental results show that DeepHalo with a true skill statistics (TSS) score of 0.907 outperforms a closely related long short-term memory network with a TSS score of 0.821. To our knowledge, this is the first time that the transformer model has been used for halo CME prediction.

Figures (8)

• Figure 1: Constructing positive and negative data samples in an AR. Data samples are collected at a cadence of 12 minutes. Each rectangular box corresponds to 1 hr and contains five data samples. (a) Gray data samples collected 24 hours before 00:00 UT, i.e., the beginning of a day, in which there is a halo CME, are labeled positive. These positive data samples are collectively referred to as a positive profile. (b) Green data samples collected 24 hours before 00:00 UT, i.e., the beginning of a day, in which there is a non-halo CME, are labeled negative. These negative data samples are collectively referred to as a negative profile.
• Figure 2: Architecture of the proposed DeepHalo model. The model contains four encoders, where each encoder is composed of a multi-head attention module and a feed-forward neural network. The multi-head attention module has eight heads, where each head is a self-attention block. The model accepts as input a profile with 120 data samples where each data sample contains 18 SHARP parameters and produces as output 1 or 0 where 1 indicates that there will be a halo CME within 24 hours and 0 indicates that there will be no halo CME within 24 hours.
• Figure 3: Analysis of feature contributions and selection of best features for the LSTM (top) and DeepHalo (bottom) models. There are 18 SHARP parameters in total. The best LSTM performance is achieved by removing 8 parameters (MEANPOT, SAVNCPP, SHRGT45, TOTUSJZ, MEANGBZ, MEANJZD, MEANALP, and AREA_ACR) and only using the remaining 10 parameters, with a TSS of 0.821, as highlighted by the gray bar in the LSTM chart at the top. The best DeepHalo performance is achieved by removing 5 parameters (MEANGBH, MEANGAM, USFLUX, MEANGBZ, and TOTUSJZ) and only using the remaining 13 parameters, with a TSS score of 0.907, as highlighted by the green bar in the DeepHalo chart at the bottom.
• Figure 4: Confusion matrices of (a) LSTM and (b) DeepHalo based on the test set used in our study.
• Figure 5: Bar graphs showing the comparison between LSTM and DeepHalo based on the test set used in our study.