Improving Physical Realism in AI Weather Forecasts
Investigates the integration of 'messy' environmental observations into AI models to enhance the physical realism and resilience of next-generation weather forecasting systems.
Messy Data, Fastnet and Resilient Forecasting: Scott Hosking at AI for Science
At the AI for Science event at the Royal Society, Scott Hosking (Alan Turing Institute) discusses how “messy” environmental observations must be made AI ready to improve forecasting and resilience, set against rising costs and impacts from extreme weather and related national security risks. He introduces the UK Met Office partnership on Fastnet, an AI weather forecasting model trained on ERA5, explains its graph neural network and multimesh design, and describes work to reduce smoothing and grid artefacts while improving physical consistency without harming skill scores. He also presents a data sparse region approach that learns directly from fragmented observations for Sub-Saharan Africa and outlines an 18 month pilot with local agencies. Hosking closes with progress toward longer range climate prediction, use of open sourced foundation models, UK supercomputing, and the importance of interdisciplinary teams and project management.