Abstract
<jats:p>Global machine learning weather prediction (MLWP) models have recently achieved impressive predictive performance, in some cases outperforming global numerical weather prediction (NWP) models at lower computational costs. Initial approaches focused on developing deterministic MLWP models. Although these models perform well on deterministic metrics, they often produce overly smooth forecasts due to the use of MSE loss and do not capture uncertainty. To address these limitations, recent efforts have shifted toward developing ensemble MLWP models. For example, ECMWF has developed AIFS ENS, a variant of its Artificial Intelligence Forecasting System (AIFS), which is capable of producing skillful global ensemble forecasts. However, most of the global MLWP ensemble models operate at coarse spatial resolution, limiting their ability to provide actionable information at fine spatial scales. This capability is crucial for national meteorological institutes such as KNMI, where high-resolution probabilistic forecasts are required for risk assessment and weather warnings. Building on recent advances in stretched-grid techniques, this study develops a pre-trained European ensemble model by combining a stretched-grid framework with ensemble generation methods. The model is trained on ERA5 and CERRA reanalysis datasets (o96 (1 degree) and 5.5 km resolution, respectively), leveraging their extensive multi-decade archives. We investigate how different CRPS loss function designs (such as multi-scale and spectral variants) affect the quality of probabilistic forecasts, with a focus on calibration, sharpness, and mitigating spatial smoothness. The resulting pre-trained model is designed to serve as a computationally efficient starting point for subsequent stretched-grid finetuning on higher-resolution reanalysis datasets, allowing meteorological institutes to streamline the development of MLWP ensemble models. This work supports the mission of KNMI to deliver skilful probabilistic forecasts for weather warnings, and contributes toward enabling operational, data-driven, high-resolution ensemble forecasting.</jats:p>