CESOC kindly invites you to a talk given by Prof Dr Julian Quinting, Professor for Energy Meteorology at the Institute for Geophysics and Meteorology and Research Scientist at the Institute of Energy Economics the Universität zu Köln, with the title:
ML-based weather forecasts for research on the sub-seasonal time scale – challenges and opportunities
Date: 3 December 2025
Time: 10:15 CEST
Location: Forschungszentrum Juelich,Building 14.14, Room BR6 (2nd floor).
For online participation, please contact info@cesoc.net
Abstract:
Weather forecasting is undergoing a radical change thanks to the extremely successful use of ML-based weather forecasting models. Models trained based on reanalysis data are achieving impressive results and show improvements over numerical weather prediction (NWP) models for a range of variables. In addition, the required computing costs are many times lower, making ML-based models attractive for generating ensemble forecasts on the sub-seasonal time scale. In this talk, I will first present different approaches for generating initial condition perturbations for ML-based ensemble forecasts in a cost-effective manner. A particular focus is on the calculation of singular vectors, which enable targeted perturbations of the initial state. In the second part of the presentation, I will introduce the concept of relaxation experiments using ML-based and NWP models to study atmospheric predictability on the sub-seasonal timescale emerging from tropical-extratropical teleconnections. The comparison shows that the loss of skill in NWP models is much more strongly influenced by epistemic uncertainties than in ML-based weather prediction models. Further analyses suggest that some of these uncertainties can be attributed to the representation of tropical waves.
Bio: https://www.ewi.uni-koeln.de/de/team/julian-quinting/