Effect of incorporating physical constraints on gray swan forecasting

Ascertain whether embedding explicit physical constraints—such as conservation laws and symmetry-equivariant architectures—into state-of-the-art AI weather/climate models enhances their ability to forecast or emulate out-of-distribution gray swan events, including Category 5 tropical cyclones.

Background

The paper assesses physical consistency of AI forecasts via gradient-wind balance and finds that both FourCastNet trained on full data and trained without strong tropical cyclones lack this balance near storm centers. Although adding physical constraints is widely proposed as a remedy, the authors emphasize that it remains unknown whether such constraints would improve gray swan forecasting.

Resolving this question would guide model design choices aimed at improving extreme-event fidelity and out-of-distribution generalization.