Physically Informed Deep Learning for Predicting Tropical Cyclone Risk in a Warming Climate

Hyunjeong Kim

Abstract: Predicting extreme tropical cyclone (TC) seasons remains challenging due to sparse data and the tendency of models to regress toward climatological averages. We propose a physics-informed CNN–Transformer trained on seven dynamically relevant ERA5 predictors, using a novel Hybrid Peak Loss that scales penalties with event magnitude to prioritize hyper-active years. Results on 45 years of data show a Mean Absolute Error (MAE) of 0.94 major hurricanes per year. Our framework demonstrates improved prediction of high-impact seasons, providing a robust baseline for climate risk assessment.

Submission Number: 7