Anfragen und Hinweise auf weitere Veranstaltungen richten Sie bitte an: outreach(at)geomar.de
09.11.2022: online Ocean Circulation and Climate Dynamics Colloquium
Dr. Baoxiang Pan, LASG/IAP/CAS, Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China: "Deep generative model for probabilistic forecast"
When? Wednesday, November 09, 2022 at 10 am
Where? ZOOM meeting room: https://geomar-de.zoom.us/j/85926085366?pwd=MnltU1F5aHVMU2VvS1FEb0NFRWxEZz09
Meeting-ID: 859 2608 5366
Kenncode: 168691
We examine the idea of training deep generative models to disentangle 1) internal climate variability, 2) model systematic bias, and 3) impact of external forcing for probabilistic forecast across weather to climate scale. In particular, we train a multi-task conditional generative model to explicitly simulate how different climate models represent the interplay between predictability signal and internal variability noise. We illustrate the choice of learning paradigm, deep net model, loss function, training details, using a case example of global seasonal forecast. Drawbacks of current practices, and promising new directions will be sketched. We close the presentation with a potentially open Q&A on data-driven climate model diagnosis, predictability study, and future CMIP experimental design.
