Modeling and Simulating Spatial Extremes by Combining Extreme Value Theory with Generative Adversarial Networks

Modeling dependencies between climate extremes is essential for risk assessment and impact studies, yet classical multivariate extreme value methods often require strong assumptions and struggle in high dimensions. Purely data-driven generative models such as GANs, on the other hand, tend to poorly represent tail dependence. We introduce evtGAN, a hybrid approach that embeds extreme value theory into generative adversarial networks to realistically model spatial dependencies in temperature and precipitation extremes. Using a stationary 2000-year climate simulation, we show that evtGAN captures tail structures and spatial patterns more accurately than standard GANs and traditional statistical models, even under limited sample sizes. This provides a practical tool for generating synthetic extreme-event fields for climate risk analysis and scenario exploration.