Physics-Guided Multimodal Learning for Flexible Tropical Cyclone Intensity Forecasting

Abstract Tropical cyclones (TCs) bring extreme precipitation and strong winds, severely disrupting transportation systems, electricity supply, and public safety. Accurate TC intensity forecasting remains challenging due to the complex and nonlinear dynamics of storm evolution. Although recent AI-based methods have shown promise, they typically rely on best-track data or geostationary satellite observations with generic fusion strategies, without explicitly modeling the consistency between physical variables and visual features. This limits their ability to accurately forecast TC intensity and capture its temporal changes. To address this problem, we propose TCFNet, a physics-guided multimodal learning framework that uses intensity variables to guide visual representation learning and jointly models TC spatial structure and temporal evolution. In addition, with flexible encoding and temporal modeling architectures, TCFNet can handle missing multimodal observations and support continuous-time forecasting, improving its adaptability in practical scenarios. Extensive experiments on two cyclone-active basins, the Western North Pacific and South Pacific, demonstrate that TCFNet outperforms prior baselines, especially in capturing intensity changes. The source code will be made publicly available.