Interpretable Deep Learning for European heatwave dynamics

Due to climate change, heatwaves are becoming more frequent and intense, with western Europe experiencing the strongest trends in the Northern Hemisphere mid-latitudes. Part of the temperature trends are caused by circulation changes. Here we deploy Deep Learning techniques to classify European heatwaves based on their atmospheric circulation and to study their associated changes over time. We train a Variational Autoencoder (VAE) to reduce the dimensionality of the heatwave samples, on large ensemble climate model simulations. We show that the VAE generalizes well to observed heatwave circulations in ERA5 reanalysis, without the need for transfer learning. We then use probabilistic clustering techniques to find physically distinct heatwave types in ERA5 reanalysis. The circulation features relevant for heatwaves in ERA5 are consistent with the climate model heatwaves. Regression analysis reveals that the Atlantic Plume type of heatwaves are becoming more frequent over time, while the Atlantic High heatwaves are becoming less frequent. Then, we introduce new and straightforward interpretability methods to study the latent space, including feature importance identification for the clustering model. We investigate which circulation features are associated with the most important nodes in the latent space and how the latent space changes over time. For example, we find that the Atlantic Low heatwave shows a deepening of the low pressure system off the Atlantic coast over time. Each heatwave type is undergoing unique changes in their circulation, highlighting the necessity to study each heatwave type separately. Our method can furthermore be used to boost specific aspects of extreme events, and we illustrate how heatwave circulation could change in the future if the current trends persist, with in some cases an intensification of features.