Greg is an atmospheric and climate scientist and Professor at the University of Washington. His research includes dynamical tests of data-driven weather models and their generalisability, particularly for weather extremes, as well as more recent work on extending predictability limits with machine-learning weather models and on how differentiable ML models can support long-window data assimilation for reanalysis. You can find examples of his recent work here: https://arxiv.org/abs/2309.10867 , https://arxiv.org/abs/2504.20238, https://arxiv.org/abs/2604.00348, and https://arxiv.org/abs/2608.11515.
Probing the Physical Limits of Extreme Events with Differentiable Weather Models
Computationally efficient and differentiable weather models enable the discovery of previously unknown extreme weather systems through gradient-based searches. In this talk, Greg will describe the search for “gray swan” extreme events that have not yet occurred but lie at the physical limits of what the atmosphere may produce.
By introducing small perturbations to ERA5, the approach identifies East Coast North American cyclones more intense than any previously observed. The physical plausibility of these simulated events is then assessed using a traditional physics-based model. Beyond dynamical studies, such events could provide synthetic training data for machine-learning weather models and inform emergency planning, infrastructure design, and insurance hazard assessment.