Statistical Learning for Climate Models

Climate models face challenges in accurately representing atmospheric circulation patterns related to extreme weather events, especially regarding regional variability. This thesis explores how Latent Dirichlet Allocation (LDA), a statistical learning method originating from natural language processing, can be adapted to evaluate the ability of climate models to represent data such as SeaLevel Pressure (SLP). LDA identifies a set of local synoptic-scale structures, physically interpretable as cyclones and anticyclones, referred to as motifs. A common basis of motifs can be used to describe reanalysis and model data so that any SLP map can be represented as a sparse combination of these motifs. The motif weights provide local information on the synoptic configuration of circulation. By analyzing the weights, we can characterize circulation patterns in both reanalysis data and models, allowing us to identify local biases, both in general data and during extreme events. A global dynamic error can be defined for each model run based on the differences between the average weights of the run and reanalysis data. This methodology was applied to four CMIP6 models. While large-scale circulation is well predicted by all models on average, higher errors are found for heatwaves and cold spells. In general, a major source of error is found to be associated with Mediterranean motifs, for all models. Additional evaluation criteria were considered: one was based on the frequency of motifs in the sparse map representation. Another one involved combining the global dynamic error with the temperature error, thus making it possible to discriminate between models. These results show the potential of LDA for model evaluation and preselection.

The defense will take place in the presence of the following jury :

• Marc BOCQUET, Professeur, CEREA, école des ponts
• Corentin HERBERT, Chargé de recherche CNRS, ENS de Lyon
• Mathieu VRAC, Directeur de recherche CNRS, LSCE
• François YVON, Directeur de recherche CNRS, ISIR, Sorbonne Université