Combining machine learning and data assimilation to estimate sea ice concentration
Combiner l'apprentissage automatique et l'assimilation de données pour estimer la concentration de glace de mer
Geer, Alan
ECMWF aims to analyse the full state of the Earth system from the atmosphere and ocean through to the land surface and cryosphere. This is intended to be achieved through closer coupling of the relevant models and the data assimilation systems that combine the model forecasts with new observations. But aspects of the Earth system, such as sea ice, snow, soil and vegetation, are hard to model from physical first principles, so in practice, the modelling can be quite empirical. Earth system modelling components are often parametrized or fitted based on a limited set of observations from experimental ground stations, and they may struggle to perform in other locations. A 'model first' approach has served us well in the atmosphere, where at least the dynamics are mostly well known: here the purpose of observations is to correct the physical trajectory of the model. But the recent explosion in machine learning for Earth system applications has shown us an 'observation first' approach. If observation-driven machine learning forecasts are starting to do better than physical forecasts, it suggests we have not been making good enough use of observations to improve our forecast models. Especially for Earth system applications where models are already partly empirical, there is clearly great potential to let the observations increasingly define these models. However, as demonstrated in this article with the example of sea ice assimilation, the best results are unlikely to come by throwing away physical models entirely, but by carefully combining known physics with empirical components.</p>
Accès à la notice sur le site du portail documentaire de Météo-France