Understanding European Heatwaves with Variational Autoencoders
Virtual poster presentation, CMIP Community Workshop 2026, Kyoto, Japan
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. In this study, we use a spatiotemporal Variational Autoencoder (VAE), an unsupervised machine learning method, to identify compact representations of year-round multivariate heatwave patterns. We extract eleven-day multivariate (e.g., circulation, humidity, temperature, geopotential height, cloud cover, stream function, and radiation) heatwave samples from ERA5 reanalysis data over the North Atlantic, based on near-surface temperature extremes in Western Europe. The VAE model is trained on heatwave samples from 1941–1990 and evaluated using 2001–2022 samples. The VAE model effectively clusters heatwave events by season, and composite maps reveal the interplay and temporal evolution between different atmospheric variables in their contributions to heatwaves over Western Europe, such as summer blocking highs and winter omega blocks. Notably, recent summer heatwaves form a distinct cluster within the latent space, pointing to a shift in atmospheric dynamics consistent with climate change. This approach offers a powerful framework to explore the spatiotemporal evolution of multivariate extremes. As a next step, we plan to apply the VAE model to CMIP simulations to evaluate how well climate models reproduce observed heatwave patterns and their associated atmospheric dynamics. This can provide a promising diagnostic to assess model skill and improve projections of heatwaves under climate change.