Evaluating Global Machine Learning Models for Tropical Cyclone Dynamics and Thermodynamics

AbstractMachine Learning Weather Prediction (MLWP) models have recently demonstrated remarkable potential to rival physics‐based Numerical Weather Prediction (NWP) models, offering global weather forecasts at a fraction of the computational cost. However, thorough evaluations are essential before considering MLWP models as replacements for NWP models. This study presents a comprehensive evaluation of four leading MLWP models—GraphCast, PanguWeather, Aurora, and FourCastNet—against observations and three state‐of‐the‐art NWP models in predicting tropical cyclones (TCs) across all tropical ocean basins. All MLWP models exhibited strong skill in forecasting TC tracks, achieving an average track error of less than 200 km at a 96‐hr forecast lead time. However, they consistently underestimated maximum sustained wind speeds compared to NWP models and observations. The low bias in TC intensity forecasts by MLWP models is linked to similar bias in their training data, along with the double penalization effect. MLWP models realistically captured the absolute vorticity patterns and their advection, demonstrating their ability to represent the dynamics underlying TC translation. They also cap