AI for Climate Science Seminar Series: AI Emulators for Weather and Climate - Promise, Pitfalls, and Paths Forward

For our October edition we are excited to announce a online lecture by Alexander Wikner from the Department of Geophysical Science of the University of Chicago.

Alexander Wikner is a postdoctoral scholar in the Department of Geophysical Sciences and a member of the Climate Extremes Theory and Data (CeTD) group. Before coming to UChicago, he received my Ph.D. and M.S. in Physics from the University of Maryland, where his thesis focused on hybridizing scientific knowledge-based models with machine learning to accurately forecast high-dimensional chaotic systems. Alexander was also a fellow in the Computation and Mathematics for Biological Networks (COMBINE) NSF NRT program at the University of Maryland. Since September 2024, he is a fellow in the Eric and Wendy Schmidt AI in Science Postdoctoral Fellowship program.

For online participation, a registration is necessary.

Please register here.

Title:
AI Emulators for Weather and Climate: Promise, Pitfalls, and Paths Forward

Abstract:
AI weather prediction models now rival physics-based systems in forecast skill, and AI climate emulators can generate thousands of years of simulation cheaply. This raises the hope of using them to answer questions currently too expensive to address with conventional GCMs; these include what-if scenarios for different perturbations and forcings, as well as sampling the tail of climate distributions. In this talk, I examine whether these models are physically faithful enough to support that hope, and what we can do when they are not. First, I show that accurate AI weather models miss the butterfly effect, so small initial-condition perturbations fail to grow as they should. This phenomenon can be traced to the coarse-graining of training data. Second, using a 92,400-year GCM ground truth, I show that AI climate emulators trained on 100 years of data produce physically realistic extremes beyond their training record, but with biased frequencies. These biases arise largely from errors in the bulk of the distribution rather than the far tail. One solution to this is AI+RES, which uses AI forecasts to guide rare-event sampling in a physics-based model. This method yields unbiased return periods for 1-in-50,000-year heat waves at roughly 100 times lower cost than direct sampling. Finally, I discuss ongoing work on a fully generative ERA5 climate emulator, designed to better capture the stochasticity and fine-scale variability that perturbation experiments depend on.