The researchers first extended this standard model into a Generalized Drag-Based Model, designed to explore potential differences in how fast and slow coronal mass ejections propagate. They then refined this generalized version into a weighted variant, which they call w-GDBM, that serves as the physical backbone of the new framework.

Because any physical model carries limitations, the team added a data-driven layer on top of the weighted model. XGBoost, a widely used machine learning algorithm, is trained to predict the residuals—the differences between the actual observed transit times and the times predicted by w-GDBM. The algorithm learns from features describing both the coronal mass ejection itself and the surrounding solar wind, allowing it to correct systematic errors that the physics model alone cannot capture.

The performance of the hybrid framework was tested across 800 random train-test splits of the data. Compared with the weighted Drag-Based Model alone, Phy-XGBoost reduced the average mean absolute error from 13.44 hours to 11.40 hours, and the average median absolute error from 11.14 hours to 8.79 hours. The researchers also compared their method with a purely data-driven XGBoost model that relied on the same features without a physical model underneath. The hybrid achieved a comparable average mean absolute error but a lower average median absolute error, suggesting that anchoring the machine learning correction in physics improves the reliability of typical predictions even when the average performance is similar.

The authors caution that these results, reported in their study accepted by Frontiers in Astronomy and Space Sciences, suggest the physics-guided machine learning framework is a promising approach to predicting coronal mass ejection arrival times rather than a finished operational forecasting system. The physical model supplies a structured first estimate grounded in the dynamics of propagation, while the machine learning component learns where that estimate tends to go wrong and adjusts it using observed conditions.

Subject of Research: A hybrid physics-guided machine learning model for predicting the arrival times of coronal mass ejections

Article Title: Constructing a new hybrid model–Phy-XGBoost to predict the arrival time of coronal mass ejections

Article References: Guan, F., Yang, Y., Xu, G., Yu, L., & Shen, F. (2026). Constructing a new hybrid model–Phy-XGBoost to predict the arrival time of coronal mass ejections. Frontiers in Astronomy and Space Sciences, 13, 1977142. https://doi.org/10.3389/fspas.2026.1977142

Image Credits: AI Generated

DOI: 10.3389/fspas.2026.1977142

Keywords: coronal mass ejections, space weather, XGBoost, machine learning, drag-based model, solar physics, forecasting, solar wind, hybrid model, transit time prediction, Constructing, hybrid