Predicting coronal mass ejection travel times using enhanced model-guided machine learning

TL;DR

This work tackles CME travel-time prediction by extending the drag-based framework to the extended drag-based model (EDBM) and embedding it within a physics-informed neural network. The two-stage approach first estimates a non-drag acceleration $a$ for Cross-wind regimes, then trains an 8-layer network with a loss that enforces EDBM-consistent travel times, enabling accurate predictions even for non-DBM-compatible events. It further introduces a six-class speed-regime classifier using multinomial logistic regression to select the applicable EDBM regime, boosting operational applicability. The results show competitive MAE around 13 hours on Cross-wind events and robust regime discrimination (average test accuracy ~0.79, TSS ~0.63), demonstrating the model’s potential for real-time, physics-guided CME forecasting and broader applicability beyond classic DBM assumptions.

Abstract

Coronal mass ejections (CMEs) are key drivers of space weather events, posing risks to both space-borne and ground-based systems. Accurate prediction of their arrival time at Earth is critical for impact mitigation. To this end, physics-informed artificial intelligence (AI) approaches have proven more effective than purely data-driven or physics-based methods, generally offering higher accuracy and better explainability than the former and lower computational cost than the latter. In this work, we propose a generalization of the physics-driven AI framework based on the classical drag-based model (DBM) by integrating the extended version of the drag-based model (EDBM). This enhancement allows us to include in the training process CME events whose interplanetary dynamics are incompatible with those assumed by the DBM. We achieve travel-time prediction accuracy comparable to state-of-the-art methods. We also perform a parametric robustness analysis, highlighting the stability of our approach under small variations in the drag coefficient. Furthermore, we propose a categorization of CMEs into speed regimes defined by the EDBM using a multiclass classification model based on logistic regression, which could be implemented in near-real-time operational space weather forecasting systems. The results show that the EDBM framework broadens the applicability of forecasting models while preserving good predictive accuracy.

Figures (4)

• Figure 1: Prediction errors on Cross-wind ($\searrow$) events across 25 runs for training (light blue box), validation (green box), and test (red box) sets. Top panels: real data. Bottom panels: augmented data. From left to right: mean absolute error (MAE), median absolute error (MedAE), and relative error. Numbers in the plots indicate mean values over the 25 realizations.
• Figure 2: Absolute-error distributions of Cross-Wind $(\searrow)$ events across 25 runs. Plots (a), (c), and (e): real events. Plots (b) and (d): augmented events. The red dashed line marks the 10-hour threshold.
• Figure 3: Confusion matrices for the multiclass classification results for training (left), validation (middle) and test (right) set of one of the 15 splits considered.
• Figure 4: Schematic flow chart of the proposed pipeline for operational CME travel-time forecasting. $\mathcal{NN}_{\text{DBM}}$ and $\mathcal{NN}_{\text{EDBM},\hat{c}}$ denote the two model-specific neural networks used to produce the predicted time $\hat{t}$, with the EDBM network conditioned on the predicted speed-regime class $\hat{c}$.