Speaker
Description
The timely and accurate prediction of Coronal Mass Ejection (CME) parameters and arrival times is a cornerstone of operational space weather forecasting. Current limitations in stereoscopic coverage hinder our ability to determine CME kinematics and geometry with confidence. To address this, our research has developed a machine learning framework trained entirely on synthetic coronagraph image sequences mimicking observations from three viewpoints. This proof-of-concept system predicts CME parameters—including speed, locations on the sun and tilt—with high accuracy, demonstrating the operational potential of data-driven multi-viewpoint approaches. Time-series coronagraph images were produced from three viewpoints (SOHO/C3, STEREO-A/COR2A, and STEREO-B/COR2B) at various cadences, with associated metadata that includes spatial and temporal parameters essential for predictive analysis. The CNN is trained using a 5-fold cross-validation method to ensure robustness. The model predicts CME longitude, latitude, and tilt with mean absolute errors (MAE) of 4.88°, 2.50°, and 13.35°, respectively, and Spearman correlation coefficients ($\rho$) of 0.93, 1.00, and 0.80. Height predictions from SOHO, STEREO-A, and STEREO-B achieve MAEs of 0.49 $R_{\odot}$, 0.42 $R_{\odot}$, and 0.42 $R_{\odot}$, with $\rho$ = 0.97 for all three. The overall model performance yields a test loss of 0.369 ± 0.009 and a test MAE of 0.136 ± 0.003.
With the upcoming launch of ESA’s Vigil mission at L5, we are approaching a transformational moment. Vigil’s continuous side-on view of the Sun and inner heliosphere will complement existing Earth-orbiting and L1 missions such as SOHO, SDO, STEREO-A, PUNCH and Aditya-L1, enabling a multi-perspective vantage essential for stereoscopic CME tracking. These coordinated viewpoints will make it possible to transition current synthetic-trained models to real data through domain adaptation technique which is a special case of transfer learning, opening the door to robust, near-real-time operational deployment.
We propose this framework as a candidate for integration into future operational pipelines, particularly for improving CME arrival time prediction at Earth. By combining Vigil’s coronagraph and EUV data with complementary datasets from L1 and Earth-orbiting spacecrafts, and embedding these into a real-time prediction model, this work lays the foundation for a new generation of machine learning based space weather tools. Our goal is to support the community in achieving full mission readiness, ensuring that from Day 1, Vigil’s observations contribute directly to operational forecasting and enhanced space weather resilience.
| Do you plan to attend in-person or online? | In-person |
|---|
