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

Area of Science:

• Atmospheric Science
• Artificial Intelligence
• Computational Physics

Background:

• Understanding AI's capability in simulating atmospheric mesoscale phenomena is crucial.
• Current AI models require rigorous evaluation against established physics-based models and reanalysis data.

Purpose of the Study:

• To compare the mesoscale kinetic energy spectra simulated by a novel AI model (Pangu) against a physics-based model (MPAS).
• To evaluate the performance of the AI model in weather forecasting skill and spectral energy distribution.

Main Methods:

• Simulated an 11-day experiment using Pangu (AI) and MPAS (physics-based) models.
• Utilized ERA5 reanalysis data as a reference for comparison.
• Employed latitude-weighted root mean square error (RMSE) and anomaly correlation coefficient (ACC) for evaluation.

Main Results:

• The AI model demonstrated superior short to medium-range weather forecasting skill (RMSE, ACC).
• The AI model underestimated mesoscale energy and failed to replicate the -5/3 spectral slope, indicating lower effective resolution.
• AI model showed stronger downscale energy flux at larger scales but weaker at smaller scales compared to the physics-based model.

Conclusions:

• The AI model (Pangu) aligns well with ERA5 at large scales, likely due to training data influence.
• Significant underestimation of mesoscale kinetic energy by the AI model compared to the physics-based model (MPAS) was observed.
• Findings highlight the AI model's limitations in resolving finer atmospheric scales and suggest caution in interpretation due to model specificity.