Authors: Syed Raza (The university of Alabama in Huntsville), Talwinder Singh (Georgia State University), Nikolai Pogorelov (The University of Alabama in Huntsville)

Coronal mass ejections (CMEs) are among the most energetic events in our solar system. Earth-directed CMEs are major drivers of geomagnetic storms, making accurate prediction of their time of arrival (TOA) at earth a key aspect of space weather forecasting. Although several models of varying capacity have been proposed over the past decades, CME TOA errors remain over 10 hours. In this study, we use a hybrid approach that combines ensemble magnetohydrodynamic simulations of CMEs, represented as flux ropes propagating through a data-driven solar wind background, with a machine learning (ML) algorithm to improve CME TOA predictions. To simulate CMEs, we use the Multiscale Fluid-Kinetic Simulation Suite (MS-FLUKSS, Pogorelov et al., 2014). We use graduated cylindrical shell (GCS) model to determine CME properties through multiple viewpoint coronagraph images, and 13 CMEs are fitted using this approach ranging from April 2010 to October 2012. To improve TOA prediction for each CME, a lasso regression ML method is used to compare the results of ensemble modeling with the data from the heliospheric imagers aboard STEREO-A and STEREO-B. We find that, using only speed ensemble members, the TOA error is reduced from 9.07 hours to 5.66 hours. Although the average TOA error decreases, not all CME TOA predictions are improved. For CMEs that do not show improvement, the errors may reflect uncertainties in other GCS parameters than speed, or the limitation of the current ML framework’s ability to capture the full complexity of CME propagation.