Temporal Tracking of a X-class flare on the Eastern-limb of the Sun with corresponding explanation maps.
The question
What does a model learn, and how can we evaluate whether its explanations are reliable?
Deep neural networks can identify complex patterns in scientific imagery, but predictive accuracy alone does not tell us whether those patterns are physically meaningful. I investigate attention mechanisms and post hoc attribution methods to examine the evidence behind solar flare predictions.
This work connects explanations across full-disk solar images and individual active regions. It also evaluates explanations using physically meaningful features and attribution-based proximity analysis, helping researchers investigate model behavior beyond a single performance score.
Related work
Explaining solar flare predictions with post hoc attribution and physical relevance
C. Pandey, J. Hong, A. Ji, M. K. Georgoulis, and B. Aydin. Poster presentation online, AGU Fall Meeting, December 2025.
View posterLarge scale evaluation of deep learning‑based explainable solar flare forecasting models with attribution‑based proximity analysis
T. Adeyeha, C. Pandey, and B. Aydin, “Large scale evaluation of deep learning‑based explainable solar flare forecasting models with attribution‑based proximity analysis,” in 2024 IEEE International Conference on Big Data (BigData), 2024, pp. 1209–1214.
Read publicationTowards reliable deep learning models for solar flare prediction
C. Pandey, R. A. Angryk, and B. Aydin. AGU poster, 2024.
View posterExplainable deep learning‑based solar flare prediction with post hoc attention for operational forecasting
C. Pandey, R. A. Angryk, M. K. Georgoulis, and B. Aydin, “Explainable deep learning‑based solar flare prediction with post hoc attention for operational forecasting,” in Discovery Science, Springer Nature Switzerland, Oct. 2023, pp. 567–581.
Read publicationExplaining full‑disk deep learning model for solar flare prediction using attribution methods
C. Pandey, R. A. Angryk, and B. Aydin, “Explaining full‑disk deep learning model for solar flare prediction using attribution methods,” in European Conference on Machine Learning and Knowledge Discovery in Databases: ADS Track (ECML PKDD), Springer Nature Switzerland, Sep. 2023, pp. 72–89.
Read publicationInsights into deep learning-based full-disk solar flare prediction with post hoc explanation and evaluation
C. Pandey, T. Adeyeha, T. Nandakumar, A. Rafal, and B. Aydin. EarthCube: A Geoscience and Cyberinfrastructure Workshop, 2023.
View poster