CNN-Based Model for 3D CME Parameter Prediction - A Proof of Concept

Accurate 3D reconstruction of coronal mass ejections (CMEs) is essential for understanding their propagation and improving space weather forecasts. In this study, we present a proof-of-concept deep learning framework that predicts seven CME parameters from synthetic multi-view coronagraph images. We generate 9,250 synthetic CME events using a 3D wireframe model, each defined by randomly sampled values for CME's longitude, latitude, tilt-angle, and 3D propagation speed. The speed is used to compute heights at three timepoints aligned with SOHO/C3 and STEREO/COR2-A/COR2-B image cadences. These values are used to render time-resolved synthetic images, serving as inputs to a convolutional neural network (CNN). The CNN is trained using five-fold cross-validation with five repeats to predict the sine and cosine of the longitude, latitude, tilt-angle, and apex heights at the three specified times. CNN achieves mean absolute errors (MAE) of 4.88°, 2.50°, and 13.35° for longitude, latitude, and tilt, and ∼0.4R⊙ for heights. The predicted heights are used to derive CME speed via height-time fitting. Performance metrics include Spearman's ρ > 0.97 for most parameters (tilt: ρ = 0.80), and R2 values of 0.99 for longitude and latitude, with heights > 0.93, and 0.61 for tilt. The average model test MAE is 0.134 ± 0.004. Applied to a real CME, the CNN's predicted longitude, latitude, and tilt align well with GCS estimates, though heights are overestimated. While challenges in using real data persist, our results on synthetic data demonstrate the effectiveness of CNN in training deep learning models for accurate 3D parameter prediction.