The GeoCME framework (Alobaid et al., 2024) showed that deep learning applied to SOHO solar images can predict, with useful skill, whether an Earth-arriving coronal mass ejection (CME) will cause a geomagnetic storm, but only for halo/partial-halo CMEs of Solar Cycle 23, conditional on arrival predictions supplied by external methods, with untested operational behavior and unexamined physical meaning of the learned image features. Follow-up work in 2026 has substantially addressed the cross-cycle, cross-instrument, and non-halo generalization question by retraining on new data (Yan et al. on Solar Dynamics Observatory (SDO) observations of Solar Cycles 24-25, though still predicting only for Earth-reaching CMEs as in GeoCME; Tang et al. on all LASCO CMEs 2001-2023 with chronological out-of-cycle testing, though labeling non-arriving events as non-geoeffective rather than conditioning on arrival). The surviving open problem is threefold and unified by one objective: (i) couple image-based geoeffectiveness prediction with CME arrival prediction into an end-to-end pipeline, starting from the general observed CME population, and quantify how arrival-prediction errors propagate into geoeffectiveness forecast skill; (ii) establish, beyond single-event case studies, which image structures drive the geoeffectiveness decision and whether they are causally connected to the in-situ magnetic field structures near 1 AU (flux-rope orientation, southward interplanetary magnetic field) that actually generate the storm; and (iii) validate the forecast under real-time operational conditions, including degraded, incomplete, and lower-cadence image streams and transfer across cycles or instruments without full retraining. Together these would determine whether image-based deep learning can yield a causally reliable, operationally deployable geoeffectiveness forecast rather than a proof-of-concept classifier.

Background

A coronal mass ejection (CME) is a large eruption of magnetized plasma from the solar corona; when an Earth-directed CME reaches the near-Earth space environment it can drive a geomagnetic storm, a global disturbance of Earth's magnetosphere that threatens satellites, power grids, and other technological infrastructure. The standard measure of storm intensity is the Dst index, an hourly index derived from low-latitude magnetometers; a storm is commonly defined by the minimum Dst falling below −50-50 nT, and a CME that produces such a storm is called geoeffective. What makes a CME geoeffective is chiefly the southward component of the interplanetary magnetic field (IMF), BzB_z, carried by the CME's magnetic flux rope (a twisted magnetic structure that, once propagating in interplanetary space, is observed as an interplanetary CME, or ICME): sustained southward BzB_z enables magnetic reconnection at the dayside magnetopause and injects energy into the magnetosphere. Unfortunately, BzB_z at Earth is not directly measurable from the Sun; remote-sensing instruments observe only projected signatures. Halo CMEs appear to surround the occulted solar disk in a coronagraph image and are the events most likely to be Earth-directed. The two-step forecasting structure of the field reflects this: an arrival model first decides whether (and when) a CME reaches Earth, and a second model then assesses storm impact. Approaches to the second step include physics-based magnetohydrodynamic propagation codes (e.g., EUHFORIA), empirical relations, and machine learning on CME catalog parameters such as speed and angular width (e.g., Pricopi et al., 2022). Alobaid et al. (2024) introduced GeoCME, which instead uses directly observed solar images as input: white-light coronagraph image sequences from LASCO C2 (LASCO is the Large Angle and Spectrometric Coronagraph), extreme-ultraviolet images at 195 Å from EIT (the Extreme-ultraviolet Imaging Telescope), and line-of-sight magnetograms from MDI (the Michelson Doppler Imager), all instruments on board the Solar and Heliospheric Observatory (SOHO), fused through an ensemble of ImageNet-pre-trained convolutional networks (transfer learning, i.e., reusing networks pre-trained on a large generic image corpus). Trained on 136 halo/partial-halo CMEs of Solar Cycle 23 (1997-2006), GeoCME predicted geoeffectiveness (Dst <−50< -50 nT) with a Matthews correlation coefficient (MCC, a balanced classification score) of 0.807 and a true skill statistic (TSS, the difference between the true-positive and false-positive rates) of 0.714, or a Brier score (BS, a probabilistic forecast score where lower is better) of 0.094 when used probabilistically. The authors note that the work relies on existing methods to decide, beforehand, which CMEs arrive at Earth. Follow-up work has since addressed the dataset's scope: Yan et al. (2026) retrained a fusion model (an ensemble combining three deep-learning image-recognition architectures: ResNet, EfficientNet, and the Vision Transformer) on 164 Earth-reaching CMEs of 2011-2024 from the Richardson-Cane list, a widely used catalog of interplanetary CMEs observed near Earth, using LASCO C2 plus extreme-ultraviolet images from AIA (the Atmospheric Imaging Assembly) and magnetograms from HMI (the Helioseismic and Magnetic Imager), both on board the Solar Dynamics Observatory (SDO), with mean TSS 0.703, and Tang et al. (2026) trained a TimeSformer video-transformer model (a deep network that applies the transformer attention architecture to sequences of image frames) with catalog parameters on 15,983 CMEs observed by SOHO/LASCO during 2001-2023 (only 116 of them geoeffective, with events that never reach Earth counted in the non-geoeffective class), evaluated with a strictly chronological split (training 2002-2014, test 2019-2023), achieving ROC-AUC 0.852 on the held-out years. Interpretability tools used in this literature include Integrated Gradients (IG), an attribution method that assigns importance scores to input pixels by integrating gradients along a path from a baseline image to the actual input, and SHAP values (SHapley Additive exPlanations, an attribution method that assigns each input feature its average marginal contribution to the prediction) for parameter-based models. What no work has yet delivered is an arrival-coupled, error-propagation-quantified, end-to-end pipeline over the general CME population, a systematic physical validation of which image features drive geoeffectiveness decisions and how they connect to in-situ magnetic structures at 1 AU, or robustness testing under real-time degraded image streams.

Problem Statement

Determine whether solar-image-based deep learning can yield a causally reliable and operationally deployable forecast of CME geoeffectiveness, by addressing three linked tasks left open after the GeoCME proof of concept (Alobaid et al., 2024) — whose reliance on existing external methods to decide which CMEs arrive at Earth, unexamined 'hidden relationships' between the solar images and geoeffectiveness, and untested 'potential for operational utilization' mark them — and its 2026 follow-ups: (1) End-to-end coupling: construct and evaluate a forecasting pipeline that starts from CME onset observations of the general observed CME population (not pre-selected Earth-arriving or halo events) and integrates arrival prediction with geoeffectiveness prediction, either as a coupled two-stage system or a joint model, together with a quantitative account of how arrival-prediction errors (missed events, false hits, timing errors) propagate into overall geoeffectiveness forecast skill, reporting skill on temporally held-out events. The end-to-end target is the unconditional probability that an onset-observed CME ultimately causes a geomagnetic storm (minimum Dst <−50< -50 nT) at Earth, with arrival prediction serving as an internal stage of the pipeline; skill conditioned on arrival may be reported alongside, but the headline skill and the error-propagation account are defined over the full population. (2) Physical interpretability and the causal link: identify, beyond single-event case studies, which image structures drive the geoeffectiveness decision of image-based models, and either establish or refute a causal connection between those structures and the in-situ magnetic field structures near 1 AU — such as flux-rope orientation and sustained southward IMF BzB_z — that physically generate the storm. Because coronal images provide only projected signatures of the three-dimensional magnetic structure, establishing or refuting this connection means demonstrating or ruling out a statistically robust, event-level correspondence between the attributed image structures and in-situ flux-rope and BzB_z measurements on temporally held-out events, under controls for solar-cycle-phase confounding; visual plausibility of attribution maps alone does not settle the question. (3) Operational robustness: characterize forecast behavior under realistic real-time data conditions, including reduced cadence, missing image frames, and saturated or otherwise degraded imagery, and determine whether cross-cycle or cross-instrument operation is achievable without full retraining or with explicit domain adaptation (statistically adjusting a model trained on one instrument or cycle so that it transfers to another). The three tasks articulate different facets of the single objective — converting demonstrated Solar-Cycle-23 classification skill into a forecast whose reliability is understood both statistically and physically under operational conditions: the coupling analysis quantifies the end-to-end forecast skill over the general population, the interpretability work examines whether that skill rests on image structures physically linked to the in-situ magnetic structures near 1 AU, and the robustness tests examine whether it survives realistic operational data conditions. All three facets are required parts of this determination rather than alternative branches: each is to be carried out and reported whatever its outcome, and the overall determination is reached only once all three have a result.

References

  1. Alobaid, K.A., Wang, J.T.L., Wang, H., Jing, J., Abduallah, Y., Wang, Z., Farooki, H., Cavus, H., Yurchyshyn, V. (2024). Prediction of Geoeffective CMEs Using SOHO Images and Deep Learning. Solar Physics 299, 159. DOI: 10.1007/s11207-024-02385-w. https://doi.org/10.1007/s11207-024-02385-w (also arXiv:2501.01011, https://arxiv.org/abs/2501.01011)
  2. Yan, Z., Wang, J.T.L., Wang, H., Lee, H., Jing, J., Xu, Y., Xu, C., Yurchyshyn, V. (2026). Deep Learning-Enabled Prediction of Geoeffective CMEs Using SOHO and SDO Observations. Solar Physics 301, 73. DOI: 10.1007/s11207-026-02665-7. https://doi.org/10.1007/s11207-026-02665-7 (also arXiv:2605.24748, https://arxiv.org/abs/2605.24748)
  3. Tang, X., Yang, X., Xu, L., Huang, X. (2026). Intelligent prediction model for geoeffectiveness of coronal mass ejections. Frontiers in Astronomy and Space Sciences 13, 1839173. DOI: 10.3389/fspas.2026.1839173. https://doi.org/10.3389/fspas.2026.1839173
  4. Fu, H., Zheng, Y., Ye, Y., Feng, X., Liu, C., Ma, H. (2021). Joint Geoeffectiveness and Arrival Time Prediction of CMEs by a Unified Deep Learning Framework. Remote Sensing 13(9), 1738. DOI: 10.3390/rs13091738. https://doi.org/10.3390/rs13091738
  5. Zhou, W., Zhou, W., Yang, Y., Cao, J., Ge, Z.-Y., Xu, T., E, P., Liu, Y., Li, S., Zhou, X., Zhao, Y., Xiong, S., Dong, Y. (2026). Coronal Mass Ejection Transit Time Prediction Based on Spatiotemporal Modeling Using 3D Convolutional Neural Networks. The Astrophysical Journal 999, 23. DOI: 10.3847/1538-4357/ae3c79. https://doi.org/10.3847/1538-4357/ae3c79
  6. Rüdisser, H.T., Davies, E.E., Amerstorfer, U.V., Möstl, C., Weiler, E., Weiss, A.J., Le Louëdec, J., Reiss, M.A., Nguyen, G. (2026). Towards a Fully Automated Pipeline for Short-Term Forecasting of In Situ Coronal Mass Ejection Magnetic Field Structure. arXiv:2602.06926. https://arxiv.org/abs/2602.06926
  7. Ye, Y., Liu, J., Hao, Y., Cui, J. (2024). Evaluating the Geoeffectiveness of Interplanetary Coronal Mass Ejections: Insights from a Support Vector Machine Approach with SHAP Value Analysis. The Astrophysical Journal 972, 52. DOI: 10.3847/1538-4357/ad61d7. https://doi.org/10.3847/1538-4357/ad61d7
  8. Pricopi, A.-C., Paraschiv, A.R., Besliu-Ionescu, D., Marginean, A.-N. (2022). Predicting the Geoeffectiveness of CMEs Using Machine Learning. The Astrophysical Journal 934, 176. DOI: 10.3847/1538-4357/ac7962. https://doi.org/10.3847/1538-4357/ac7962 (also arXiv:2206.11472, https://arxiv.org/abs/2206.11472)