MS5A – Data-Driven Regional Weather Modeling: Towards Trustworthy Convection-Resolving Forecasts
Session Chair
Event Type
Minisymposium
Climate, Weather, and Earth Sciences
Physics
Computational Methods and Applied Mathematics
TimeWednesday, July 114:00 – 16:00 CEST
LocationBldg. 6 – 002
DescriptionOrganizer(s): Oliver Fuhrer (MeteoSwiss, ETH Zurich), and Laure Raynaud (Météo-France)
Data-driven weather prediction has advanced rapidly in recent years, with machine-learning-based models now complementing traditional numerical weather prediction. While global data-driven models have demonstrated impressive skill, extending these approaches to regional, convection-resolving forecasting introduces new scientific and computational challenges. At kilometer and sub-kilometer scales, models must represent complex physical processes, integrate high-frequency observations, and provide trustworthy uncertainty estimates, particularly for extremes. This minisymposium focuses on recent progress and open challenges in data-driven regional weather modeling. Topics include generative diffusion models for convective-scale downscaling, graph-based neural network architectures for high-resolution domains, training strategies that improve generalization across scales and regions, and operational perspectives from national meteorological services. The session emphasizes scientific trustworthiness, evaluation, and physical consistency, and discusses how these requirements interact with high-performance computing workflows and model design. By bringing together experts from academia and operational forecasting, the minisymposium provides a forum to assess the state of the art and explore pathways toward reliable, convection-resolving data-driven forecasts, with relevance to a wide range of computational science domains.
Data-driven weather prediction has advanced rapidly in recent years, with machine-learning-based models now complementing traditional numerical weather prediction. While global data-driven models have demonstrated impressive skill, extending these approaches to regional, convection-resolving forecasting introduces new scientific and computational challenges. At kilometer and sub-kilometer scales, models must represent complex physical processes, integrate high-frequency observations, and provide trustworthy uncertainty estimates, particularly for extremes. This minisymposium focuses on recent progress and open challenges in data-driven regional weather modeling. Topics include generative diffusion models for convective-scale downscaling, graph-based neural network architectures for high-resolution domains, training strategies that improve generalization across scales and regions, and operational perspectives from national meteorological services. The session emphasizes scientific trustworthiness, evaluation, and physical consistency, and discusses how these requirements interact with high-performance computing workflows and model design. By bringing together experts from academia and operational forecasting, the minisymposium provides a forum to assess the state of the art and explore pathways toward reliable, convection-resolving data-driven forecasts, with relevance to a wide range of computational science domains.
Presentations
