Forecasting the arrival time of Earth-directed coronal mass ejections (CMEs) via physics-based simulations is an essential but challenging task in space weather research due to the complexity of the underlying physics and limited remote and in-situ observations of these events. Data assimilation (DA) techniques can assist in constraining free model parameters and reduce the uncertainty in subsequent model predictions. In this study, we show that CME simulations conducted with the Space Weather Modeling Framework (SWMF) can be assimilated with SOHO LASCO white-light (WL) observations and solar wind observations at L1 prior to the CME eruption to improve the prediction of CME arrival time. The L1 observations are used to constrain the model of the solar wind background into which the CME is launched. Average speed of CME shock front over propagation angles are extracted from both synthetic WL images from the Alfv\’en Wave Solar atmosphere Model (AWSoM) and the WL observations. We observe a strong rank correlation between the average WL speed and CME arrival time, with the Spearman’s rank correlation coefficients larger than 0.90 for three events occurring during different phases of the solar cycle. This enables us to develop a Bayesian framework to filter ensemble simulations using WL observations, which is found to reduce the mean absolute error of CME arrival time prediction from about 13.4$\pm$3.8 hours to 5.1$\pm$3.0 hours. The results show the potential of assimilating readily available L1 and WL observations within hours of the CME eruption to construct optimal ensembles of Sun-to-Earth CME simulations.

Modeling the impact of space weather events such as coronal mass ejections (CMEs) is crucial to protecting critical infrastructure. The Space Weather Modeling Framework (SWMF) is a state-of-the-art framework that offers full Sun-to-Earth simulations by computing the background solar wind, CME propagation and magnetospheric impact. However, reliable long-term predictions of CME events require uncertainty quantification (UQ) and data assimilation (DA). We take the first steps by performing global sensitivity analysis (GSA) and UQ for background solar wind simulations produced by the Alfvén Wave Solar atmosphere Model (AWSoM) for two Carrington rotations: CR2152 (solar maximum) and CR2208 (solar minimum). We conduct GSA by computing Sobol indices that quantify contributions from model parameter uncertainty to the variance of solar wind speed and density at 1 au, both crucial quantities for CME propagation and strength. Sobol indices also allow us to rank and retain only the most important parameters, which aids in the construction of smaller ensembles for the reduced-dimension parameter space. We present an efficient procedure for computing the Sobol indices using polynomial chaos expansion (PCE) surrogates and space-filling designs. The PCEs further enable inexpensive forward UQ. Overall, we identify three important model parameters: the multiplicative factor applied to the magnetogram, Poynting flux per magnetic field strength constant used at the inner boundary, and the coefficient of the perpendicular correlation length in the turbulent cascade model in AWSoM.

One major challenge of solar flare prediction with machine learning methods is the scarcity of large flares. This issue of low positive sample size is even more severe for data observed in the relatively weak Solar Cycle 24, for example, the SHARPs data product. This partly hampers the successful application of deep learning methods, especially those dealing with high-dimensional spatial and/or temporal data. By joining SHARPs with Space-Weather MDI Active Region Patches (SMARPs), a new data product derived from observations in Solar Cycle 23, we are able to obtain a fused dataset with nearly tripled positive samples. We evaluated two deep learning methods, LSTM and CNN, using the selected parameter sequences and image snapshots in the fused dataset. Experiment results show that the two models trained on the fused dataset achieve better or equivalent test set performance than those trained on a single solar cycle. In addition, we demonstrate the improvement of the performance of the stacking ensemble that combines LSTM and CNN. We provided interpretation to CNN using modern visual attribution methods in computer vision. The results show that CNN is able to identify flare-related signatures in magnetograms.

In this work, we develop gradient boosting machines (GBMs) for forecasting the SYM-H index multiple hours ahead using different combinations of solar wind and interplanetary magnetic field (IMF) parameters, derived parameters, and past SYM-H values. Using Shapley Additive Explanation (SHAP) values to quantify the contributions from each input to predictions of the SYM-H index from GBMs, we show that our predictions are consistent with physical understanding while also providing insight into the complex relationship between the solar wind and Earth’s ring current. In particular, we found that feature contributions vary depending on the storm phase. We also perform a direct comparison between GBMs and neural networks presented in prior publications for forecasting the SYM-H index by training, validating, and testing them on the same data. We find that the GBMs have a comparable root mean squared error as the best published black-box neural network schemes.

We use machine learning methods to predict whether an active region (AR) which produces flares will lead to a solar energetic particle (SEP) event using Space-Weather Michelson Doppler Imager (MDI) Active Region Patches (SMARPs). This new data product is derived from maps of the solar surface magnetic field taken by the Michelson Doppler Imager (MDI) aboard the Solar and Heliospheric Observatory (SOHO). We survey the SMARP active regions associated with flares that appear on the solar disk between June 5, 1996 and August 14, 2010, label those that produced SEPs as positive and the rest as negative. The AR SMARP features that correspond to each flare are used to train two different types of machine learning methods, the support vector machines (SVMs) and the regression models. The results show that the SMARP data can predict whether a flare will lead to an SEP with accuracy (ACC) {less than or equal to}0.72{plus minus}0.12 while allowing for a competitive leading time of 55.3{plus minus}28.6 minutes for forecasting the SEP events.

Many current research efforts undertake the solar flare classification task using the Space-weather HMI Active Region Patch (SHARP) parameters as the predictors. The SHARP parameters are scalar quantities based on spatial average or integration of physical quantities derived from the vector magnetic field, which loses information of the two-dimensional spatial distribution of the field and related quantities. In this paper, we construct two new sets of spatial features to expand the feature set used for the flare classification task. The first set uses the idea of topological data analysis to summarize the geometric information of the distributions of various SHARP parameters. The second set utilizes tools coming from spatial statistics to analyze the vertical magnetic field component Br and summarize its spatial variations and clustering patterns. All features are constructed within regions near the polarity inversion lines (PILs) and classification performances using the new features are compared against the SHARP parameters (also along PIL). We found that using the new features can improve the skill of the flare classification model and new features tend to have higher feature importance, especially the spatial statistics features. This potentially suggests that even using a single channel Br, instead of all SHARP parameters, one can still derive strongly predictive features for flare classification.

In this talk, we present our machine learning efforts, which show great promise towards early predictions of solar flare events. (1) We present a data pre-processing pipeline that is built to extract useful data from multiple sources – Geostationary Operational Environmental Satellites (GOES) and Solar Dynamics Observatory (SDO)/Helioseismic and Magnetic Imager (HMI) and SDO/Atmospheric Imaging Assembly (AIA) – to prepare inputs for machine learning algorithms. (2) For our strong/weak flare classification model, case studies show a significant increase in the prediction score around 20 hours before strong solar flare events, which implies that early precursors appear at least 20 hours prior to the peak of a flare event. (3) We develop a mixed Long Short Term Memory (LSTM) regression model to predict the maximum solar flare intensity within a 24-hour time window. (4) Our ongoing and future work will also be briefly mentioned.