Towards a Climate Counterfactual Autoencoder

Separating forced climate change from internal climate variability is a fundamental challenge in climate attribution. In the emerging field of extreme event attribution, climate models are used to produce so-called `storyline simulations' of climate extreme events: those events are simulated under a constrained atmospheric circulation in factual (=present-day) and counterfactual (=without anthropogenic forcing) climate conditions in order to attribute the thermodynamic effects of anthropogenic climate change. However, traditional approaches for producing such circulation‑conditioned counterfactuals are computationally costly, cannot be directly transferred to observations, and cannot be easily transferred to other climate conditions than the ones simulated. Here we show that deep learning offers large potential to generate highly versatile climate counterfactuals: we use a Variational Autoencoder to predict counterfactual European winter temperatures by providing only the global mean warming level (i.e., the background climate) and the atmospheric circulation state as inputs. The results are benchmarked against traditional nudged-circulation climate model simulations. The deep learning based counterfactuals are shown to perform extremely well and can be applied in any background climate state, thus providing a versatile climate counterfactual generator. Future work could target counterfactual climate states based on observed weather states. Accurate climate counterfactuals could strongly support climate adaptation and communication efforts.