Seminar Speaker-Turco Lecturer: Professor Maria Molina, University of Maryland

Are we ready to ask "what-if" with AI weather emulators?

AI weather models have made substantial progress in medium-range forecasting, with some models outperforming ECMWF’s ensemble prediction system on selected verification targets. This progress motivates a further question: can AI weather models support “what-if” studies that probe how an event’s evolution responds to changes in its initial conditions or surrounding environment? We investigate this question using gradient-based guided diffusion with DeepMind's GenCast, steering its sampling toward specified tropical cyclone characteristics and analyzing how the resulting forecast trajectories respond. GenCast is a graph-based conditional diffusion model with an encoder, graph-transformer processor, and decoder. We find that guidance toward a broader initial wind field can alter Sandy’s (2012) westward track. For Andrew (1992), guidance toward storm-size estimates produces a track resembling the National Hurricane Center forecast. Guiding Irma’s (2017) evolution also changes the subsequent simulated track and intensity of Maria (2017). Together, these experiments illustrate how guided diffusion can probe sensitivities in AI-generated weather scenarios and raise a key question: how can we distinguish physically meaningful sensitivities from responses specific to the model and guidance?

https://illinois.zoom.us/j/87247933909?pwd=Yg9a5afZrUDAodWJBhap3HvxzMVVeo.1

Meeting ID: 872 4793 3909     Pass Code: 202496