Can AI models predict out-of-distribution gray swan weather extremes?

AI models are transforming weather and climate prediction and increasingly outperforming physics-based models. Here, I present results from several tests and state-of-the-art AI models demonstrating that these models fail to predict the rarest, yet most impactful, weather extremes - especially those so rare that they are entirely absent from the training data (so-called gray swans). I will then show how combining AI models with mathematical tools from rare-event sampling and using active learning can address this problem.

Bio

Pedram Hassanzadeh is an Associate Professor at the University of Chicago's Department of Geophysical Sciences and Committee on Computational and Applied Math. He is the faculty director of the AI for Climate Initiative (AICE) and co-director of the Human-centered Weather Forecasts Initiative. He received his MA and PhD from UC Berkeley and was a Ziff Environmental Fellow at Harvard University. His research is at the intersection of climate change, scientific machine learning, computational and applied math, extreme weather, and turbulence physics.