Learning Extremes with evtGAN

Extreme environmental events such as floods and heatwaves are difficult to model because traditional approaches require either computationally expensive climate simulations or strong parametric assumptions that fail in high-dimensional, non-stationary settings. This talk introduces evtGAN, a generative model that merges extreme value theory with GANs by separating marginal tail modelling from dependence learning. Using copula transformations and GEV-based normalization, evtGAN accurately reproduces spatial tail behaviour and extrapolates beyond the training range using as few as 30 annual maxima, offering a fast, flexible, and accurate emulator for extremes.