Understanding European Heatwaves with Variational Autoencoders

Abstract. Understanding the dynamics of heatwaves is critical for accurate climate risk assessment. Traditional definitions, based solely on surface temperature thresholds, often overlook the complex, multivariate nature of heatwaves. This study uses a spatiotemporal Variational Autoencoder (VAE), an unsupervised machine learning method, to identify compact representations of multivariate, year-round heatwave patterns. Focusing on key atmospheric variables (e.g., circulation, humidity, temperature, geopotential height, cloud cover, stream function, and radiation), we extract eleven-day heatwave samples from ERA5 reanalysis data over the North Atlantic, centered on near-surface temperature extremes in Western Europe. The VAE was trained on data from 1941–1990 and evaluated using 2001–2022 samples, and effectively clustered heatwave events by season, revealing known dynamical regimes such as summer blocking highs and winter omega blocks. The VAE model captures the interplay and temporal evolution between different atmospheric variables in their contributions to heatwaves over Western Europe. Notably, recent summer heatwaves form a distinct cluster within the latent space, pointing to a shift in atmospheric dynamics consistent with climate change. Composite anomaly maps further show coherent pre-onset patterns across variables. These results demonstrate the potential of VAEs to uncover meaningful structure in complex heatwave dynamics from data, and promise advances in understanding heatwaves.