Group seminar on 25. May, 14:15 CET

Machine-Learning-Assisted Numerical Modeling of the Atmosphere?

The advances in atmospheric modeling and data assimilation techniques, alongside with the increase of computer power and the number of observations available for assimilation led to a “quiet revolution of NWP”. The incorporation of machine learning (ML) techniques into the NWP process offers perhaps the most promising approach to extend the gains in forecast accuracy by extracting further information from the observations. This talk presents a hybrid modeling approach, in which ML is used to periodically and interactively correct the state vector of the spatiotemporally evolving numerical model solution. Results are shown for an implementation of the approach on a low-resolution primitive equation model. They indicate that the forecast performance of the hybrid model is better than that of the host numerical model, a pure ML-based model, or a model that uses linear regression rather than ML for the correction of the numerical forecasts. They also show that the hybrid model maintains a more realistic atmospheric balance between the mass and momentum fields than the host model at all forecast times. The potentials of the approach for climate modeling is demonstrated by the results of a decade-long climate run with the hybrid model.