Estimating Coronal Mass Ejection Arrival with Ensemble Physics-Driven Machine Learning: The May 2024 Superstorm Case

Abstract The extreme G5 geomagnetic storm recorded in May 2024 was the consequence of a very fast and energetic Coronal Mass Ejection (CME) that cannibalized a sequence of multiple eruptions previously emitted by the Sun. This dramatic plasma cloud reached Earth in less than two days, thus demonstrating that resilient measures to safeguard in-orbit and on-Earth technological assets are very critical when such extreme events occur. Using remote-sensing coronal observations and in-situ solar-wind measurements, this study shows that an innovative ensemble physics-driven machine-learning method would have been able to anticipate the CME travel time for this single out-of-sample stress-test event with a point prediction error of 1 minute from the observed arrival, while the ensemble spread, reflecting training-related epistemic uncertainty, indicates a characteristic uncertainty of approximately 3 hours. In addition, a sensitivity analysis was performed to assess the robustness of the approach, evaluating how uncertainties in the input parameters propagate into the predicted travel time. Results showed strong robustness, with the mean predicted travel time remaining within a few minutes of the actual arrival time, and the mean absolute error around 3 hours when input uncertainties were taken into account. Furthermore, by benchmarking our method against purely data-driven and classical drag-based models, we demonstrate that, under the extreme conditions of the May 2024 superstorm, the proposed hybrid approach yields significantly superior accuracy.