Mankind faces severe climate change-induced problems in the near future. How well we can face these problems strongly depends on a solid scientific base for the political and societal decisions to be made (IPCC, 2021a). A very large fraction of this scientific base is built on numerical modelling (IPCC, 2021b). Climate projections such as those of the Coupled Model Intercomparison Project Phase 6 (CMIP6, Eyring et al., 2016) form the cornerstones of our knowledge of the times to come and aid the upcoming decision making process far beyond the physical base they provide (cf., migration, diseases, pests, crops). However, these climate projections do have a major drawback. Especially with coupled Earth System Models (ESM) as the CMIP-models, each free-running model simulation represents only a single realization out of the literally unlimited trajectories possible. Studying specific questions such as “how would the German Ahrtal flooding change in a 2-degree warmer climate”  or “how would the 2021 European heat wave (Fig. 1a) evolve in 2028”  is nearly impossible with such a classic setup. The reason is that the chances of a big heat wave occurring in 2028 in one of the models are very slim, let alone one that closely resembles the 2021 heat wave. Even general questions such as how a 3-degree warmer world with 20% increased CO2 would look like still pose a problem today.

CMILE compresses large climate data sets by employing variational auto-encoders (VAE). In our VAE, nonlinear climate inter-relations are stored. It then becomes possible to explore and exploit these inter-relations in the latent space of the VAE.

By analyzing the VAE’s latent space dimension and vectors can be identified that represent climate events or trends.  Even the temporal evolution of a single latent space dimension stores patterns that represent whole climate-specific features as seasonality, climate trends or El Niño (Fig. 1).