Predicting arrival time for CMEs: Machine learning and ensemble methods

Coronal mass ejections (CMEs) are arguably one of the most violent explosions in our solar system. CMEs are also one of the most important drivers for space weather. CMEs can have direct adverse effects on several human activities. Reliable and fast prediction of the CMEs arrival time is crucial to minimize such damage from a CME. We present a new pipeline combining machine learning (ML) with a physical drag-based model of CME propagation to predict the arrival time of the CME. We evaluate both standard ML approaches and a combination of ML + probabilistic drag based model (PDBM, Napoletano et al. 2018). More than 200 previously observed geo-effective partial-/full-halo CMEs make up the database for this study (with information extracted from the Richardson & Cane 2010 catalogue, the CDAW data centre CME list, the LASCO coronagraphic images, and the HEK database - Hurlburt et al. 2010). The P-DBM provides us with a reduced computation time, which is promising for space weather forecasts. We analyzed and compared various machine learning algorithms to identify the best performing algorithm for this database of the CMEs. We also examine the relative importance of various features such as mass, CME propagation speed, and height above the solar limb of the observed CMEs in the prediction of the arrival time. The model is able to accurately predict the arrival times of the CMEs with a mean square error of about 9 hours.  We also explore the differences in prediction from ML models and emblem prediction method namely P-DBM model.