Algorithmic Hallucinations of Near Surface Winds: Statistical Downscaling with Generative AI

Providing small-scale information about weather and climate is challenging, especially for variables strongly controlled by physical processes that are unresolved by low-resolution (LR) simulations. In this talk, I will present some of my research using emerging machine learning methods from the fields of image super-resolution and deep learning for statistically downscaling (i.e. image upsampling) near-surface wind patterns. Specifically, I train Generative Adversarial Networks (GANs), which are conditioned on LR inputs from a global climate model, to generate high-resolution (HR) surface winds that emulate those simulated in an HR climate model (native 4-km grid spacing). Since explicitly simulating weather and climate at HR is prohibitively computationally expensive for operational use, using GANs instead can greatly improve the accessibility of fine-scale climate information to communities, thus making them an exciting tool for climate practitioners.

Bio:
Nic Annau is a Physical Scientist with Environment and Climate Change Canada in the Climate Research Division. His research interests include generative AI, computer vision, software development, and statistical downscaling.

Watch a recording of the talk.