Developing a new high-resolution weather prediction model using AI

Recent advancements in AI-based Weather Prediction (AIWP) are causing a paradigm shift in global weather forecasting.

Recent data-driven models, in particular, Google DeepMind's GraphCast, have demonstrated that AI-based models can match or even outperform the best global numerical weather prediction model, namely ECMWF’s Integrated Forecasting System, while significantly reducing computation time. Building on these developments, ECMWF has made substantial progress in AIWP, with their current AI forecasting system, surpassing GraphCast in accuracy for e.g. temperature (see Figure 1). However, for operational use at KNMI, these forecasts need to be generated at much higher resolution (~ 2 km) than that of AIFS (currently ~31 km). Furthermore, previous research has highlighted the limitations of AIWP models in predicting precipitation and extreme weather events (see Figure 2), necessitating targeted improvements for these events as well.

The goals of this project are, therefore, three-fold. Firstly, we will tackle the need of localized higher resolution forecasting by following a stretched-grid approach (see Figure 3 for an example) within the AIFS model, similar to Nipen et al. (2024). This will allow for higher-resolution forecasts in a specific region while maintaining a lower-resolution on the rest of the globe. Secondly, since weather forecasts are inherently uncertain, we will develop ensemble forecasts based on the stretched-grid paradigm, following the recent advancements in probabilistic forecasting from Google DeepMind, ECMWF and Met Norway. Lastly, we will investigate the usage of different loss functions for training AIWP models that facilitate the forecasting of extreme events while, at the same time, having a limited impact on the models’ overall accuracy.