AI for Climate Science seminar series: How much climate data do we need? Machine learning approaches for evaluating high-resolution climate models

Join us for the next AI for Climate Science Seminar!

In our June edition we will hear from Climate Dynamics and Modeling experts from the University of Vienna -  Aiko Voigt and Maximilian Meindl.

Aiko Voigt, born in 1980 in Berlin, studied physics at Humboldt University of Berlin and Paul Sabatier University in Toulouse. After completing his doctorate at the Max Planck Institute for Meteorology, he conducted research in Paris, New York, and Karlsruhe. Since 2021, he has been Professor of Climate Sciences at the University of Vienna.

Maximilian Meindl, born in 1997 in Vienna, studied meteorology at the University of Vienna. Since graduating in 2024, he has been working as a research assistant at the Department of Meteorology and Geophysics at the University of Vienna. In addition to his research in climate modeling, he is a member of klimaszenarien.at, an initiative focused on developing new climate scenarios for Austria. 

For online participation, a registration is necessary.

Please register here.

Title:
How much climate data do we need? Machine learning approaches for evaluating high-resolution climate models

Abstract:
The emergence of kilometer-scale climate models represents a major step forward in the simulation of atmospheric processes, allowing many phenomena to be represented explicitly rather than through parameterization. However, these simulations also produce enormous data volumes and are often limited to relatively short time periods due to their high computational cost. As a result, traditional climate model evaluation methods based on long climatological averages become increasingly difficult to apply and may provide limited insight when only short segments of high-frequency output are available.
In the first part of this talk, we focus on a machine learning (ML) based approach for evaluating climate models on regional scales using short periods of daily near-surface temperature fields. A convolutional neural network (CNN) is trained to distinguish spatial temperature patterns from simulations of the EURO-CORDEX ensemble and global kilometer-scale models. Applying the trained CNN to observation datasets provides an alternative evaluation metric based on similarities between model-generated and observed spatial patterns.
In a second step, we apply Layerwise Relevance Propagation (LRP) as an explainable AI method to interpret the CNN’s decisions. This allows us to identify which spatial regions and structures are most relevant for classification and to relate them to physically meaningful features.
Finally, we show ongoing work exploring whether the climatological bias structure of climate models can be inferred directly from short samples of daily near-surface temperature fields using ML-based spatial feature extraction.