Capabilities, Limitations and Validation of Statistical Earth System Models
DescriptionThe last year has seen tremendous progress on deep learning based forecasting models trained on reanalysis data. An extension of these to Earth system models (ESMs) and to training on observational data for obtaining predictions that are not bound by the reanalysis process (and hence classical models) is an exciting frontier in the coming years. I will first discuss how machine learning-based models are conceptually different from classical ESMs in that they are statistical representations of the dynamical processes modeled by p( y|x ). Some capabilities that statistical models can provide that are very difficult to impossible to achieve with classical ESMs will be showcased and potential fundamental limitations of such models will be discussed. Based on this, I will examine novel challenges for validating machine learning-based Earth system models, in particular when trained on observational data.
TimeWednesday, June 2815:30 - 16:00 CEST
Climate, Weather and Earth Sciences