Probabilistic weather forecasts provide essential information about forecast uncertainty, which is especially important for high-impact weather such as severe wind gusts. To obtain reliable predictions, numerical weather prediction (NWP) model output is typically statistically post-processed, but most current approaches are limited to single locations and ignore spatial structure. This project develops spatial post-processing techniques that leverage modern machine learning, particularly generative models, to produce calibrated and spatially consistent ensemble forecasts directly on the NWP model grid. By incorporating spatial information through deep neural networks and optimizing mathematically principled multivariate proper scoring rules, the project aims to enhance forecasts of wind gusts and temperature in collaboration with the German Weather Service.