Forecasting Extreme Space Weather Events with Physics-Driven Machine Learning: CME Arrival Prediction for the May 2024 Superstorm

The G5 geomagnetic superstorm of May 2024 represents one of the most extreme space weather events in the space era and provides a critical testbed for assessing our preparedness for future severe storms. The event was driven by an exceptionally fast and energetic Coronal Mass Ejection (CME) that resulted from the cannibalization of multiple preceding eruptions, producing a complex plasma structure that reached Earth in less than two days. Such short warning times underscore the urgent need for robust and accurate forecasting frameworks to protect space- and ground-based technological infrastructures. In this study, we investigate the predictability of the May 2024 superstorm using an ensemble, physics-driven machine learning approach that combines remote-sensing coronal observations with in-situ solar wind measurements. Our results show that this hybrid framework would have successfully predicted the CME Sun–Earth travel time with high accuracy, achieving a timing precision of the order of one minute and an uncertainty of approximately three hours. A sensitivity analysis was conducted to assess the robustness of the model against uncertainties in the input parameters. The analysis demonstrates strong stability of the forecasting framework, with the mean predicted arrival time remaining within a few minutes of the observed value and a mean absolute error of about three hours when realistic input uncertainties are considered. Furthermore, benchmarking against classical drag-based models and purely data-driven approaches reveals that the proposed hybrid method significantly outperforms existing techniques during this extreme event.These results highlight the potential of physics-driven machine learning as a key component of next-generation space weather forecasting systems. Finally, this contribution also discusses possible future improvements in extreme CME propagation and addresses the challenges related to the validation of forecasting methodologies, with a particular focus on assessing prediction skill and robustness.