WS 2026

Prof. Sebastian Lerch
Titel: Uncertainty quantification for data-driven weather prediction: From probabilistic forecasts to fair model comparisons 

Abstract: 

Artificial intelligence (AI)-based data-driven weather forecasting models have experienced rapid progress over the last years. Recent studies achieve impressive results and demonstrate substantial improvements over state-of-the-art physics-based numerical weather prediction models across a range of variables and evaluation metrics. Beyond improved predictions, the main advantages of data-driven weather models are their substantially lower computational costs and the faster generation of forecasts, once a model has been trained. However, most efforts in data-driven weather forecasting have been limited to deterministic, point-valued predictions, making it impossible to quantify forecast uncertainties, which is crucial in research and for optimal decision making in applications. Moreover, fair comparisons of data-driven and physics-based models are challenging, with data-driven models operating by optimizing a training loss and using the prespecified loss function for this purpose, unlike physics-based models. In this talk, I will give an overview of recent work on uncertainty quantification methods to
(1) generate probabilistic forecasts from deterministic data-driven weather models, and to
(2) enable fair and meaningful comparisons of data-driven and physics-based deterministic models.

Federico Lot
Titel: Numerical Aspects of Multiscale Approximation

In many practical applications the need arises to construct high-fidelity models from discretely given
data. Kernel methods are a popular choice as they can cope with unstructured point clouds. The main
drawback is, however, that usually a dense linear system with high condition number has to be solved.
Kernel-based multiscale methods have proven to provide a remedy to those limitations. These
methods are characterized by an appropriate scaling of a kernel and a hierarchical organization of the
data. The resulting approximant is known to be accurate as well, but its computation is substantially
more stable.
In this talk we will discuss both recent analytical understandings and novel approaches to the
implementation of the kernel based multiscale method.
This work is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation)– Projektnummer 452806809.
Details can be found at the preprint: arXiv:2503.04914.