On the Extrapolation of Generative Adversarial Networks for Downscaling Precipitation Extremes in Warmer Climates

While deep-learning downscaling algorithms can generate fine-scale climate projections cost-effectively, it is unclear how effectively they extrapolate to unobserved climates. We assess the extrapolation capabilities of a deterministic Convolutional Neural Network baseline and a Generative Adversarial Network (GAN) built with this baseline, trained to predict daily precipitation simulated by a Regional Climate Model (RCM) over New Zealand. Both approaches emulate future changes in annual mean precipitation well, when trained on historical data, though training on a future climate improves performance. For extreme precipitation (99.5th percentile), RCM simulations predict a robust end-of-century increase with future warming (similar to 5.8%/degrees $\mathit{{}<^>{\circ}}$C on average from five simulations). When trained on a future climate, GANs capture 97% of the warming-driven increase in extreme precipitation compared to 65% in a deterministic baseline. Even GANs trained historically capture 77% of this increase. Overall, GANs offer better generalization for downscaling extremes, which is important in applications relying on historical data.