Enriching Operational High-Resolution Ensemble Forecasts with StyleGAN-2

• DOI: 10.1175/aies-d-24-0058.1
• Publication Date: 2025-03-24T15:51:58Z
• Authors: Clément Brochet; Gabriel Moldovan; Julien Rabault; Cyril Regan; Laure Raynaud

Abstract The main approach to account for uncertainty in numerical weather prediction (NWP) is the use of Ensemble Prediction Systems (EPS), which run several simulations parallel. However, huge cost these systems, especially high-resolution modeling, severely limits their implementation, and particular number ensemble members. This size constraint can have significant impact on accuracy robustness distributions sampling, as well utility EPS low-predictability situations. study introduces an innovative hybrid forecasting system, combining strengths NWP generative deep learning. proposed method leverages properties StyleGAN latent space create a large new members at reduced cost. developed using dataset forecasts produced by kilometer-scale AROME-EPS operational Météo-France. We first illustrate capability GAN generate realistic samples with reasonable physical consistency. It then shown that GAN-enriched few hundreds able significantly enhance sampling distribution tails, way consistent obtained very AROME-EPS. Probabilistic verification scores are also promising, although they point out some weaknesses current set up. Existing bottlenecks avenues improvements discussed. Deep models potential transform significantly. In this paper, we initiate exploration into vast possibilities methods offer.