Posters 1

[Poster 1-09] Graph Neural Networks for Nested Weather Prediction from Global to Regional Scales

*Yuki Yasuda (1. Japan Agency for Marine-Earth Science and Technology (Japan))

Digital twins of the ocean require accurate atmospheric prediction because ocean circulation is driven in part by surface wind stress and air-sea heat fluxes. Deep learning models have advanced rapidly in both atmospheric and ocean prediction, making them important components of Earth system digital twins. In the atmosphere, accurate prediction of extreme weather often requires kilometer-scale grids capable of resolving mesoscale structures. However, global prediction at such resolution remains computationally demanding. Limited-area models (LAMs) offer a practical alternative by restricting high-resolution prediction to a regional domain and using lateral boundary conditions from a coarser global model. However, the use of deep learning in LAMs and their integration with global models remain underexplored.

This study develops a nested prediction system that combines a global model with a LAM over Japan. The global model is a message-passing graph neural network trained on the global reanalysis dataset (JRA-3Q). It operates at a horizontal grid spacing of approximately 125 km, and its predictions provide lateral boundary conditions for the LAM during both training and prediction. The LAM is based on a graph attention network and is trained on the regional reanalysis dataset (RRJ-Conv). It operates at a grid spacing of approximately 5 km. By nesting the two models, the system downscales coarse global predictions into high-resolution predictions over Japan. We evaluate prediction accuracy at both global and regional scales using standard verification metrics, and we examine extreme weather over Japan through case studies. In the poster session, we will discuss the potential and limitations of this data-driven nested approach to regional weather prediction.