Forecasting Coronal Mass Ejections Using Transformers: A Comparative Study of Flare‐Associated Versus Flare‐Independent Approaches

Abstract Forecasting coronal mass ejections (CMEs) remains a challenge, and no reliable, accurate forecasting method has yet been developed. Knowing that CMEs can occur in association with flares, we compare two approaches: (a) forecasting CMEs by coupling to flare forecasting, and (b) forecasting CMEs independently of flare occurrence. To do this, we train three transformer‐based models that use sequences of HMI‐SHARP keywords: (a) a flare model that predicts flares of class , (b) a flare‐CME model that predicts CMEs associated with flares, and (c) an independent CME model that predicts CMEs regardless of flare activity. When evaluated independently, the flare, flare‐CME, and CME models achieve True Skill Score, Matthews Correlation Coefficient, and Average Precision (AP) of , , and , respectively. When the flare and flare‐CME forecasts are chained and evaluated only during the 24 hr preceding a flare, the coupled model outperforms the stand‐alone CME model with Bayesian confidence and a mean (95% Highest Density Interval (HDI) [0.08, 0.27]). However, in terms of AP, the independent CME model outperforms the coupled model with Bayesian confidence and a mean (95% HDI [0.09, 0.21]). All metrics vary with the fraction of active regions that produce events, irrespective of class imbalance, so this ratio should be reported alongside the scores. We conclude that the community needs to agree on what constitutes a “best” model, and that further progress may require knowledge of the state of the solar corona.