Exploring the differences in atmospheric mesoscale kinetic energy spectra between AI based and physics based models

It is an urgent need to understand the ability of current artificial intelligence (AI) models in simulating atmospheric mesoscale aspects. This paper compares mesoscale kinetic energy spectra from an 11-day experiment simulated by a novel AI-based model (Pangu) and a physics-based model (MPAS), using ERA5 reanalysis as a reference. Based on the commonly used evaluation metrics of latitude weighted root mean square error (RMSE) and anomaly correlation coefficient (ACC), the AI-based model has better short to medium-range weather forecasting skill compared to the physics-based model. However, the AI-based model cannot replicate the mesoscale - 5/3 spectral slope and underestimates the mesoscale energy at wavelength smaller than 1000 km. As altitude increases and scale decreases, the deviation of the AI-based model from the reanalysis significantly increases. These features prove that the AI-based model has the lower effective resolution compared to the physics-based model with the close nominal resolution. Compared to the physics-based simulations, AI-based model has stronger downscale energy flux at larger mesoscales, which is dominated by divergent kinetic energy flux. But it rapidly becomes the weakest at smaller mesoscales. The diagnosed vertical velocity of AI-based model and its related budget terms are closest to those of the reanalysis at large scales. Overall, the AI-based model Pangu shows closer agreement with ERA5 at large scales, likely due to its use of the latter as training data, but significantly underestimates mesoscale kinetic energy compared to the physics-based model MPAS. Note that these findings are specific to the models and configurations used and should be interpreted with caution.