Integrating Machine Learning with MHD modeling and Heliospheric Imaging for Improved CME Arrival Time Predictions

Coronal mass ejections (CMEs) are the primary cause of severe space weather which negatively impacts many of our space-related activities. The arrival time prediction of CMEs is an area of active research. Various techniques of varying complexity have been proposed to predict the arrival of CMEs, yet even the most sophisticated models have struggled to reduce the mean absolute error below 12 hours. In our study, we propose a new approach for predicting CME arrival time that employs magnetohydrodynamic simulations of a data-informed flux rope-based CME model, positioned within a data-driven solar wind background. By examining 6 CMEs, we managed to reduce the mean absolute error in arrival time prediction to 8 hours. We further enhanced the precision of our arrival time predictions through ensemble modeling and by comparing the ensemble results with heliospheric imager data from STEREO A and B, by creating synthetic J-maps based on our simulations. We incorporated a machine learning method known as lasso regression for this cross-comparison. By doing so, we were able to bring down our mean absolute error to 4.1 hours, indicating a notable advance in the prediction of CME arrival time. Furthermore, when we incorporated neural networks, we achieved a mean absolute error of 5.1 hours using heliospheric imager data solely from STEREO A. Consequently, our study underlines the critical role of combining machine learning techniques with other models to enhance the accuracy of space weather predictions.