2026 workshop program
A Foundation Model Approach to Solar Filament Detection
The work
Abstract
The Surya foundation model provides a powerful, generalized representation of solar observations that enables efficient and scalable downstream heliophysics analysis. In this project, we use Surya as the primary input for a scientifically relevant downstream task: classifying solar filaments and characterizing their physical properties. We developed a multi-input neural network that combines local HMI magnetic-field data with global Surya embeddings to analyze solar magnetic activity. The pipeline preprocesses and normalizes the data, calculates Polarity Inversion Line (PIL) strength as a physics-based target, and trains the model for both PIL prediction and filament detection. The model was tested on unseen solar observations using regression metrics, while filament-detection experiments are still in progress. Our results show that the model successfully learned global magnetic and PIL structures, achieving strong PIL prediction performance with (R² ≈ 0.77) and correlation coefficient (r ≈ 0.88). This work demonstrates that foundation-model embeddings can help improve machine-learning approaches for understanding complex solar magnetic behavior.