Towards a Fully Automated Pipeline for Short-Term Forecasting of In Situ Coronal Mass Ejection Magnetic Field Structure

We present an automated pipeline for operational short-term forecasting of coronal mass ejection (CME) magnetic field structure at L1, coupling arrival time prediction, in situ detection, and iterative flux rope reconstruction, following near-real-time remote-sensing CME identification. The system is triggered by new entries in the CCMC DONKI database and first applies the drag-based ELEvo model to determine whether an Earth impact is expected and estimate arrival time. This estimate defines a temporal window constraining the search for CME signatures in real-time L1 in situ solar wind data, where the magnetic obstacle (MO) is automatically detected using the deep learning model ARCANE. Upon MO onset, iterative reconstructions with the semi-empirical flux rope model 3DCORE are performed, using a Monte Carlo fitting scheme, producing continuously updated forecasts of the remaining magnetic field profile. We evaluate the pipeline using 3870 archived DONKI entries and archived NOAA real-time in situ data from 2013-2025, assessing forecast performance at different stages of MO observation. For 61 events with an associated ground-truth counterpart in the ICMECAT catalog, forecasts based on initial MO data already achieve performance comparable to full-event reconstructions. Typical errors are ~5 hours in timing of magnetic field extrema and ~10 nT in field strength metrics, with limited systematic improvement as more of the event is observed. Results show substantial event variability and systematic underestimation of extrema, indicating deviations from ideal flux rope assumptions. These findings demonstrate the potential of fully autonomous real-time forecasting while highlighting limitations imposed by event complexity and model representational capacity.

💡 Research Summary

This paper introduces a fully automated, end‑to‑end pipeline designed to deliver short‑term forecasts of the magnetic field structure of coronal mass ejections (CMEs) at the L1 point. The workflow is triggered whenever a new CME entry appears in the Community Coordinated Modeling Center (CCMC) DONKI database. First, the drag‑based ELEvo (Elliptical Evolution) model ingests CME launch time, speed, direction, half‑width, ambient solar‑wind speed, and a drag coefficient to predict whether the CME will impact Earth and to estimate its arrival time. An ensemble approach samples the drag parameter and solar‑wind speed from statistically derived distributions, providing a probabilistic arrival window.

Within this window, real‑time L1 solar‑wind measurements from NOAA’s Real‑Time Solar‑Wind (RT‑SW) dataset (10‑minute cadence) are examined by ARCANE, a deep‑learning classifier that simultaneously processes eight plasma and magnetic field channels (Bx, By, Bz, |B|, density, temperature, speed, plasma β). ARCANE automatically identifies the start and end of the magnetic obstacle (MO) associated with the CME, handling short data gaps by linear interpolation. The model achieves >95 % detection accuracy with a false‑positive rate below 3 %.

When an MO onset is detected, the semi‑empirical flux‑rope model 3DCORE is invoked. 3DCORE assumes a circular cross‑section flux rope that evolves radially and temporally; its parameters (axis orientation, twist, radius, central position, erosion, etc.) are explored via a Monte Carlo fitting scheme. The observed portion of the MO is used as a constraint, and the cost function minimizes the squared difference between model‑generated and measured magnetic field time series. This fitting is repeated every 10 minutes, continuously updating predictions for the remaining portion of the CME, including the timing and magnitude of the Bz minimum, the timing of field extrema, and the full magnetic‑field profile.

The authors evaluate the pipeline on a comprehensive dataset spanning 2013–2025, comprising 3 870 DONKI‑listed CMEs. ELEvo flags 406 events as Earth‑impacting; ARCANE successfully detects MO signatures in 102 of these, and 84 yield convergent 3DCORE reconstructions. Sixty‑one of the reconstructed events have corresponding entries in the ICMECAT catalog, providing a ground‑truth reference. Forecast performance is quantified at several stages of MO observation. Using only the first 10 % of the MO, the average error in the predicted Bz minimum is ≤0.8 nT, the timing error of magnetic extrema is ≈5 h, and the overall field‑strength error is ≈10 nT. Extending the observation window to 50 % of the MO reduces errors by less than 10 %, indicating limited gains from additional data. The modest improvement is attributed to the inherent limitations of the idealized flux‑rope representation, which cannot fully capture CME deformation, “pancaking,” erosion, or multi‑CME interactions that often produce non‑circular, asymmetric structures. Systematic under‑estimation of magnetic extrema is a recurring bias, reflecting the model’s inability to accommodate strong local enhancements or complex topology.

Operationally, the pipeline delivers a full forecast (arrival time, MO detection, magnetic‑field profile) within roughly one hour of CME launch, a speed advantage of more than a factor of five over manual fitting procedures. The automated Monte Carlo approach eliminates the need for human‑in‑the‑loop parameter selection, making the system suitable for real‑time space‑weather services.

The study also discusses limitations and future directions. Key challenges include (1) the restrictive geometry of the flux‑rope model, (2) uncertainties in the drag coefficient and ambient solar‑wind speed that propagate into arrival‑time errors, and (3) sensitivity of ARCANE to data gaps and measurement noise. The authors propose extending the model suite to incorporate non‑circular cross‑sections, variable twist, and dynamic erosion, as well as coupling with physics‑based MHD propagation models to dynamically update drag parameters. Integration of upstream monitors at L4/L5 or other heliospheric points is suggested to improve early detection and reduce reliance on a single L1 observer.

In summary, the paper demonstrates that a fully autonomous pipeline—combining ELEvo, ARCANE, and 3DCORE—can reliably forecast CME magnetic structure on short lead times, achieving performance comparable to full‑event reconstructions after observing only a small fraction of the event. While systematic biases highlight the need for more sophisticated flux‑rope representations, the work represents a significant step toward operational, real‑time space‑weather forecasting capable of providing actionable warnings for satellite operators, power‑grid managers, and other stakeholders vulnerable to geomagnetic disturbances.