Atmospheric circulation representation in CMIP6 models for extreme temperature events using Latent Dirichlet Allocation

Climate models aim at representing as closely as possible the statistical properties of the climate components, including extreme events on which fine-tuning data may be less available. This is a fundamental requirement to correctly project changes in their dynamics due to anthropogenic forcing. In order to estimate how much models can be trusted, we must evaluate how closely they match observations. We need algorithms capable of selecting, processing and evaluating relevant dynamical features of the climate components. This has to be reiterated efficiently for large datasets such as those issued from the Coupled Model Intercomparison Project 6 (CMIP6). In this work, we use Latent Dirichlet Allocation (LDA), a statistical soft clustering method initially designed for natural language processing, to extract synoptic patterns from sea-level pressure data. We propose to use LDA as a tool for extracting new information from model data and evaluating their performance.

LDA allows for learning a basis of decomposition of maps into objects called "motifs". From the ERA5 sea-level pressure data, the method robustly extracts a basis of motifs that are interpretable objects at synoptic scale, i.e. cyclones or anticyclones associated to locations. Pressure data can be projected onto this basis, yielding motif weights that contain local information about the large-scale atmospheric circulation. LDA decomposition is efficient and sparse: most of the information of a given map is contained in few motifs. It is therefore possible to decompose any map in a limited number of easy-to-interpret synoptic objects. This allows for a variety of new angles for statistical analysis.

The weights statistics can be used to characterize general and extreme event-specific dynamics in reanalysis and model data. By comparing the statistics obtained from reanalysis data with those obtained from a selection of CMIP6 models, we can quantify errors on each localized circulation pattern and identify model-specific and model-independent errors. We find that, thought models make higher errors on extreme events cases such as heatwaves and cold spells, on average, large-scale circulation patterns are well predicted by the models.

By projecting model simulations of several future scenarios, we can also measure changes in time of synoptic patterns predicted by the models. These changes can be general, season-specific or extreme event-specific. We find several predicted changes in motif weights statistics, some of which are consistent with currently observed changes in synoptic configuration statistics. By comparing the amplitude of the changes in different shared socio-economic pathways, we can link some of these changes to anthropogenic climate forcing, at least according to the models.