FuXi-TC: A generative framework integrating deep learning and physics-based models for improved tropical cyclone forecasts

Tropical cyclones (TCs) are among the most devastating natural hazards, yet their intensity remains notoriously difficult to predict. Numerical weather prediction (NWP) models are constrained by both computational demands and intrinsic predictability, while state-of-the-art deep learning-based weather forecasting models tend to underestimate TC intensity due to biases in reanalysis-based training data. Here, we present FuXi-TC, a diffusion-based generative forecasting framework that combines the track prediction skill of the FuXi model with the intensity representation of NWP simulations. Trained on outputs from the Weather Research and Forecasting (WRF) model driven by FuXi background fields, FuXi-TC corrects systematic biases and reconstructs realistic TC structures without retraining global models. Across 21 TCs in 2024, FuXi-TC substantially reduces root mean square error in 5-day intensity forecasts relative to both FuXi and ERA5, achieving skill comparable to the high-resolution deterministic forecasts of the European Centre for Medium-Range Weather Forecasts (ECMWF). Moreover, FuXi-TC delivers forecasts within seconds, over three orders of magnitude faster than WRF, enabling real-time operational use and large ensembles. These results demonstrate the value of integrating generative diffusion models with physics-informed datasets to advance TC prediction, offering both improved accuracy and transformative computational efficiency for disaster preparedness. More broadly, they highlight the potential of using physics-based simulations to construct new training datasets that further enhance deep learning based weather forecasting.