Speaker
Description
Forecasting astrophysical flares in blazars presents a unique challenge due to their irregular temporal dynamics and strong variability. While deep neural networks have shown promise for modeling such complex time series, their predictions often lack alignment with established physical knowledge, limiting trust and interpretability. In this work, we propose a domain-informed deep learning regularization framework that explicitly integrates astrophysical priors into the training process. Specifically, we incorporate two complementary regularization terms based on Grad-CAM explanations of model attention. First, an attention alignment loss encourages the model to focus on flux peaks, reflecting domain knowledge that flare events correspond to sharp increases in photon flux. Second, a temporal smoothness loss enforces consistency in attribution across adjacent time steps, capturing the physical duration of flare events. The combined objective function guides the model not only to minimize prediction error but also to generate explanations consistent with astrophysical priors. This approach improves both predictive accuracy and interpretability, paving the way toward more trustworthy machine learning models for flare forecasting in high-energy astrophysics.
