Exploring Physical Mechanisms for Superior Tropical Cyclone Track Forecasting in AI‐Based Ensemble Weather Prediction Models

Abstract Tropical cyclones (TCs) are highly destructive and inherently uncertain systems. Ensemble forecasting is critical for quantifying these uncertainties, yet traditional systems are constrained by high computational costs and limited representation of atmospheric nonlinearity. AI‐based ensemble forecasting has emerged as a promising paradigm and shows encouraging performance in TC forecasting, yet the underlying physical mechanisms remain to be explored. Here, we take FuXi‐ENS, which generates ensemble forecasts using a flow‐dependent learnable perturbation scheme, as a representative case and compare it against ECMWF‐ENS for all 90 global TCs in 2018. Results show that FuXi‐ENS produces smaller ensemble mean TC track errors than ECMWF‐ENS, which is physically consistent with its more accurate representation of large‐scale circulation and deep layer mean steering flow (DLMSF) governing TC motion. Nevertheless, FuXi‐ENS still exhibits underdispersive ensemble behavior. To investigate the structural characteristics of ensemble perturbations, we further employ moist turbulent energy (MTE) as a diagnostic metric to characterize the spatial distribution of perturbation energy, revealing that MTE in FuXi‐ENS is more tightly concentrated around the TC warm core, whereas ECMWF‐ENS exhibits a broader and more dispersed distribution. These findings deepen understanding of the physical processes underlying AI‐based TC track forecasting and offer valuable insights for improving the reliability of AI‐based ensemble weather prediction systems.