TC-Next Improves Cyclone Forecasting with Multimodal AI

Accurate forecasting of tropical cyclone track and intensity is critical for disaster preparedness and mitigation. Traditional rule-based trackers often have limitations, and even advanced AI models can struggle with generalization. This paper introduces TropicalCycloneNext (TC-Next), a novel multimodal deep learning model designed to enhance cyclone forecasting. TC-Next leverages two distinct data modalities: forecast fields of atmospheric kinematic and thermodynamic variables from foundation weather models (like GraphCast, Pangu-Weather, IFS HRES) and GridSat infrared satellite imagery. The model was trained exclusively on GraphCast forecasts over the Western Pacific. Despite this focused training, TC-Next exhibits remarkable zero-shot generalization capabilities. When applied without retraining to forecast fields from other models (Pangu-Weather, IFS HRES, WeatherNext Cyclones), TC-Next consistently outperforms conventional rule-based trackers like TempestExtremes, reducing track error by 15-44% and intensity error by a factor of 3-6. It also achieves lower or comparable error rates compared to specialized direct trackers from other advanced models. Ablation studies confirm that the multimodal approach, integrating satellite imagery, is key to these performance improvements, especially for tracking errors and longer-lead intensity predictions.