HAPpy Hour Seminar : Improving AI weather prediction models using global mass and energy conservation schemes

Abstract : Artificial intelligence (AI) models can learn how to make weather forecasts directly from historical data. They are known as AI weather prediction models (AIWP). These models are purely data-driven, and their forecasts do not obey global mass and energy conservation. This study provides numerical schemes to help AIWP models obey conservation laws. We compared the results from two AIWP model runs; one is purely data-driven, and the other implemented our conservation schemes. The latter conserves mass and energy better; their forecast errors are also reduced. This shows that our conservation schemes can guide AIWP models in producing more skillful and physically consistent forecasts. This also means that AIWP models can potentially receive more benefits by incorporating other physics-based relationships in the future.

At the end of this seminar, I will also provide a brief update about the regional AI weather prediction model that is under active development at RAL/NCAR as part of my PS1 work.

View recordings of previous meetings here

View the schedule (and sign up to give a talk!) here