Assessing the ability of a stretched-grid deep learning weather prediction model to learn the physics of windstorm Poly

Abstract

Weather forecasting has long relied on Numerical Weather Prediction (NWP) models, which simulate weather by solving the governing fluid equations. The recent emergence of global DeepLearning Weather Prediction (DLWP) models potentially marked the beginning of a new era in weather prediction. DLWP models have demonstrated performance on par with, or even slightly better than, the global Integrated Forecasting System (IFS) model, developed by the European Centre for Medium-Range Forecasting (ECMWF). Therefore, ECMWF has developed also their own DLWP model, called Artificial Intelligence/Integrated Forecasting System (AIFS). MET Norway, in collaboration with ECMWF, has recently overcome the resolution limitation of the initial global DLWP models by developing Bris, a stretched-grid version of AIFS with high-resolution (2.5 km) on a limited domain. The generation of Bris opened the door to the potential operational use of these weather models, prompting KNMI to develop its own stretched-grid DLWP model. As a national meteorological service, KNMI is particularly concerned with improving the prediction of extreme weather events, posing significant risks to public safety and infrastructure. Initial evaluation using standard skill metrics revealed that the Bris model and other DLWP models perform poorly in predicting extreme events. This underperformance is a result of using the Mean Squared Error as a loss function and may stem also from limited training under such conditions or from an insufficient ability to capture the complex dynamics driving these extreme events. To explore this issue further, the present study focuses on the direct comparison of Bris and the control run of the operational MetCoOp Ensemble Prediction System (MEPS) in forecasting the violent extratropical cyclone Poly, which hit the Netherlands on 5 July 2023. Following an approach similar to Bonavita (2024), who analyzed a lower-resolution global DLWP model, we assess whether Bris accurately represents deviations from key atmospheric balances. Our findings show that, despite its high resolution, Bris struggles to capture the mesoscale dynamics of the event. Additionally, Bris disrupts significantly some of the examined atmospheric balances, due to the presence of a fine-scale noise in its output fields, which leads to incorrect and unrealistic spatial gradients. These limitations raise critical questions about how to improve AI-based models to better represent extreme events and how to ensure physically consistent predictions.