Integrating Machine Learning and Solar Physics for Enhanced Prediction of CME Arrival Times and Near-Sun Solar Wind Conditions

The timely and precise prediction of coronal mass ejection (CME) arrival times and the characterization of near-Sun solar wind conditions are essential for space weather forecasting and planetary sciences. We develop a novel deep-learning framework that integrates imaging observations and physical parameters to predict CME arrival times with improved accuracy. Using time-series data from synchronized solar white-light and EUV observations of 156 geoeffective CME events (2000–2020), we train two models: Model A, a convolutional neural network (CNN) regression model, and Model B, an enhanced version incorporating 11 key physical parameters of CMEs and background solar wind. Model B achieves a minimum mean absolute error (MAE) of 5.12 hours, a 33% improvement over Model A. This demonstrates the value of combining observational and physical data in forecasting CME arrival times.

In addition, we explore the use of GONG/ADAPT magnetograms with a U-Net-based architecture to model solar wind conditions at 0.1 AU. The training labels are derived from the COCONUT coronal model, which offers a potential acceleration in generating initial driving conditions for heliophysical models like ICARUS. While preliminary, this approach highlights a pathway to streamline the modeling of near-Sun solar wind environments, further supporting interplanetary CME propagation studies.

Our results underscore the potential of machine learning when synergized with solar physics to advance predictions critical to heliophysics and planetary sciences.