go through: Professor Sarah Dance (Professor of Data Assimilation)
Last week, 25Th In February 2025, our colleagues at ECMWF (European Medium-Scale Weather Forecasting Center) adopted learning-based deep learning Global weather forecasting system, known as AIFSenter operational production, and operate with a physically based numerical weather forecasting system. AIF performs better than state-of-the-art physics models, is used for multiple accuracy and runs faster in computation. However, both systems are currently necessary to initialize and best estimate the atmospheric state created using a process called Data assimilation.
Tens of millions of atmospheric observations are used every day in weather forecasts, but observation alone cannot describe the weather in all parts of the world. To get a complete picture, we use Data assimilation To combine observations with information from a physics-based computer model, consider our confidence in each data source. The final analysis is used as a starting point for both physical and machine-based weather forecasts. It is also important to note that using data assimilation with historical observations, reanalysis is a key component of the training data of machine learning prediction systems.
Figure: Schematic diagram from the World Meteorological Organization's global observation system
In recent decades, stable improvements in global numerical prediction accuracy have been driven by increased data assimilation methods and increased observations (e.g., Bauer et al., 2015). However, new observations are expensive, with new satellites costing billions of pounds. Therefore, investment in such systems is evidence-based. To inform these financial decisions, data assimilation scientists have traditionally conducted quantitative experiments to estimate the impact of new observations on improving predictions through physically-based numerical weather forecasting systems (e.g., HU et al., 2025).
In the age of machine learning prediction, the connection between each observation type and prediction accuracy is less transparent. It is unclear how to measure the impact of a particular observation on machine learning training, nor is it clear how the initialization of machine learning models is. This is a key issue to ensure continued improvement in forecast accuracy and better adaptation to dangerous weather events.
refer to
Ball (P.nature 52547-55 (2015). https://doi.org/10.1038/nature14956
Hu, G., Dance, SL, Fowler, A., Simonin, D., Waller, J., Auligne, T. ,wait. (2025) The value of methods for evaluating observations in data assimilation and numerical weather forecasts of convection. Royal Meteorological Society Quarterly https://doi.org/10.1002/qj.4933
